[{"role":"user","text":"Help me create the content for this new skill file: /home/pastilhas/projects/officer.dev/data/skills/mlxaudio/SKILL.md\nfrom https://github.com/Blaizzy/mlx-audio"},{"role":"assistant","text":"\n\nLet me start by reading the current (empty) SKILL.md file and fetching information about the mlx-audio project."},{"role":"tool","toolName":"Read","toolInput":{"file_path":"/home/pastilhas/projects/officer.dev/data/skills/mlxaudio/SKILL.md"},"toolUseId":"toolu_01MXmpTWFrzx5Wm2fUTBEqcQ","output":"     1→---\n     2→name: mlx.audio\n     3→description: \n     4→---\n     5→","isError":false},{"role":"tool","toolName":"Read","toolInput":{"file_path":"/home/pastilhas/projects/officer.dev/data/skills/whisper.cpp/SKILL.md"},"toolUseId":"toolu_01RfdTasfwALeQhE26dkENTD","output":"     1→---\n     2→name: whisper.cpp\n     3→description: Transcribe audio files to text using whisper.cpp. Use when the user wants to transcribe audio, convert speech to text, or extract text from an audio/video file.\n     4→---\n     5→\n     6→# Whisper.cpp\n     7→\n     8→API reference for the whisper.cpp HTTP server running at `http://macmini:8178`.\n     9→\n    10→whisper.cpp is a C/C++ port of OpenAI's Whisper speech recognition model. The server accepts audio files via HTTP and returns transcriptions in various formats.\n    11→\n    12→## Server\n    13→\n    14→- **Base URL:** `http://macmini:8178`\n    15→- **No authentication required**\n    16→\n    17→## Endpoints\n    18→\n    19→### POST /inference\n    20→\n    21→Transcribes an audio file. Accepts `multipart/form-data`.\n    22→\n    23→#### Example\n    24→\n    25→```bash\n    26→curl -s http://macmini:8178/inference \\\n    27→  -F file=\"@/path/to/audio.mp3\" \\\n    28→  -F temperature=\"0.0\" \\\n    29→  -F temperature_inc=\"0.2\" \\\n    30→  -F response_format=\"json\"\n    31→```\n    32→\n    33→#### Parameters\n    34→\n    35→##### File (required)\n    36→\n    37→| Parameter | Type | Description |\n    38→|-----------|------|-------------|\n    39→| `file` | file | Audio file to transcribe. Accepts at least WAV and MP3. |\n    40→\n    41→##### Response Format\n    42→\n    43→| Parameter | Type | Default | Description |\n    44→|-----------|------|---------|-------------|\n    45→| `response_format` | string | `json` | Output format: `json`, `verbose_json` (or `vjson`), `text`, `srt`, `vtt` |\n    46→\n    47→##### Language\n    48→\n    49→| Parameter | Type | Default | Description |\n    50→|-----------|------|---------|-------------|\n    51→| `language` | string | `en` | Spoken language code (e.g. `en`, `pt`, `es`, `fr`). Use `auto` for auto-detection. |\n    52→| `detect_language` | bool | `false` | Exit after detecting the language (no transcription). |\n    53→| `translate` | bool | `false` | Translate from source language to English. |\n    54→\n    55→##### Decoding\n    56→\n    57→| Parameter | Type | Default | Description |\n    58→|-----------|------|---------|-------------|\n    59→| `temperature` | float | `0.0` | Sampling temperature. `0.0` is deterministic. |\n    60→| `temperature_inc` | float | `0.2` | Temperature increment on fallback attempts. |\n    61→| `best_of` | int | `2` | Number of candidate decodings to keep. |\n    62→| `beam_size` | int | `-1` | Beam search size. `-1` disables beam search. |\n    63→| `entropy_thold` | float | `2.40` | Entropy threshold — decoder fails and retries if exceeded. |\n    64→| `logprob_thold` | float | `-1.00` | Log probability threshold for decoder failure. |\n    65→| `no_fallback` | bool | `false` | Disable temperature fallback on decode failure. |\n    66→\n    67→##### Segmentation\n    68→\n    69→| Parameter | Type | Default | Description |\n    70→|-----------|------|---------|-------------|\n    71→| `max_len` | int | `0` | Maximum segment length in characters. `0` for unlimited. |\n    72→| `max_context` | int | `-1` | Maximum text context tokens to store. `-1` for unlimited. |\n    73→| `split_on_word` | bool | `false` | Split segments at word boundaries instead of token boundaries. |\n    74→| `no_timestamps` | bool | `false` | Suppress timestamps in output. |\n    75→| `word_thold` | float | `0.01` | Word timestamp probability threshold. |\n    76→\n    77→##### Audio Processing\n    78→\n    79→| Parameter | Type | Default | Description |\n    80→|-----------|------|---------|-------------|\n    81→| `offset_t` | int | `0` | Time offset in milliseconds — skip this much audio from the start. |\n    82→| `offset_n` | int | `0` | Segment index offset. |\n    83→| `duration` | int | `0` | Duration of audio to process in milliseconds. `0` for all. |\n    84→| `audio_ctx` | int | `0` | Audio context size. `0` for all. |\n    85→\n    86→##### Speaker Diarization\n    87→\n    88→| Parameter | Type | Default | Description |\n    89→|-----------|------|---------|-------------|\n    90→| `diarize` | bool | `false` | Enable speaker diarization (requires stereo audio). |\n    91→| `tinydiarize` | bool | `false` | Enable tinydiarize (requires a tdrz model). |\n    92→\n    93→##### Voice Activity Detection (VAD)\n    94→\n    95→| Parameter | Type | Default | Description |\n    96→|-----------|------|---------|-------------|\n    97→| `vad` | bool | `false` | Enable VAD preprocessing. |\n    98→| `vad_threshold` | float | `0.50` | Speech confidence threshold (0.0–1.0). |\n    99→| `vad_min_speech_duration_ms` | int | `250` | Minimum speech segment duration in ms. |\n   100→| `vad_min_silence_duration_ms` | int | `100` | Minimum silence duration to split segments. |\n   101→| `vad_max_speech_duration_s` | float | `FLT_MAX` | Auto-split segments longer than this (seconds). |\n   102→| `vad_speech_pad_ms` | int | `30` | Padding added around speech segments (ms). |\n   103→| `vad_samples_overlap` | float | `0.10` | Overlap between segments (seconds). |\n   104→\n   105→##### Other\n   106→\n   107→| Parameter | Type | Default | Description |\n   108→|-----------|------|---------|-------------|\n   109→| `prompt` | string | `\"\"` | Initial prompt to condition the model (e.g. for vocabulary hints). |\n   110→| `suppress_nst` | bool | `false` | Suppress non-speech tokens. |\n   111→| `no_context` | bool | `false` | Do not use previous audio context for subsequent segments. |\n   112→| `debug_mode` | bool | `false` | Enable debug output. |\n   113→\n   114→#### Response Formats\n   115→\n   116→##### `json` (default)\n   117→\n   118→Minimal JSON with just the transcribed text.\n   119→\n   120→```json\n   121→{\"text\": \"The transcribed content goes here.\"}\n   122→```\n   123→\n   124→##### `verbose_json` (or `vjson`)\n   125→\n   126→Extended JSON including task type, language, audio duration, per-segment timestamps, token-level timing, confidence scores, and language probability distribution.\n   127→\n   128→##### `text`\n   129→\n   130→Plain text transcription. Includes speaker labels if diarization is enabled.\n   131→\n   132→##### `srt`\n   133→\n   134→SubRip subtitle format with sequential numbering, `HH:MM:SS,mmm` timestamps, and text content.\n   135→\n   136→```\n   137→1\n   138→00:00:00,000 --> 00:00:03,500\n   139→The transcribed content goes here.\n   140→```\n   141→\n   142→##### `vtt`\n   143→\n   144→WebVTT subtitle format with `WEBVTT` header and `HH:MM:SS.mmm` timestamps.\n   145→\n   146→```\n   147→WEBVTT\n   148→\n   149→00:00:00.000 --> 00:00:03.500\n   150→The transcribed content goes here.\n   151→```\n   152→\n   153→### POST /load\n   154→\n   155→Loads a different model file on the server at runtime.\n   156→\n   157→```bash\n   158→curl -s http://macmini:8178/load \\\n   159→  -F model=\"/path/to/model.bin\"\n   160→```\n   161→\n   162→| Parameter | Type | Description |\n   163→|-----------|------|-------------|\n   164→| `model` | string | Path to the model file on the server. |\n   165→\n   166→## Supported Audio Formats\n   167→\n   168→The server accepts at least WAV (16-bit PCM) and MP3 files directly. If the server was started with `--convert`, it can use ffmpeg to handle additional formats (ogg, flac, m4a, etc.).\n   169→\n   170→## Source\n   171→\n   172→- Repository: https://github.com/ggml-org/whisper.cpp\n   173→- Server docs: https://github.com/ggml-org/whisper.cpp/blob/master/examples/server/README.md\n   174→","isError":false},{"role":"tool","toolName":"Read","toolInput":{"file_path":"/home/pastilhas/projects/officer.dev/data/skills/ffmpeg/SKILL.md"},"toolUseId":"toolu_0164FTK5nf9HT8m3GArB4jgy","output":"     1→---\n     2→name: ffmpeg\n     3→description: Process audio and video files using ffmpeg/ffprobe. Use when the user wants to convert, transcode, trim, merge, extract, resize, compress, or analyze multimedia files.\n     4→---\n     5→\n     6→# FFmpeg\n     7→\n     8→CLI reference for FFmpeg v8.x — a complete, cross-platform solution for recording, converting, and streaming audio and video.\n     9→\n    10→Official docs: https://www.ffmpeg.org/documentation.html\n    11→\n    12→## Tools\n    13→\n    14→| Tool | Purpose |\n    15→|------|---------|\n    16→| `ffmpeg` | Transcode, convert, filter, mux/demux multimedia |\n    17→| `ffprobe` | Analyze and inspect multimedia streams |\n    18→| `ffplay` | Play multimedia files (interactive) |\n    19→\n    20→---\n    21→\n    22→## ffmpeg\n    23→\n    24→### Synopsis\n    25→\n    26→```\n    27→ffmpeg [global_options] {[input_options] -i input_url} ... {[output_options] output_url} ...\n    28→```\n    29→\n    30→Options before `-i` apply to the input; options before the output URL apply to the output.\n    31→\n    32→### Global Options\n    33→\n    34→| Flag | Description |\n    35→|------|-------------|\n    36→| `-y` | Overwrite output files without asking |\n    37→| `-n` | Do not overwrite; exit if output exists |\n    38→| `-hide_banner` | Suppress copyright/build info banner |\n    39→| `-loglevel level` | Set log level: `quiet`, `error`, `warning`, `info` (default), `verbose`, `debug` |\n    40→| `-stats` | Print encoding progress/statistics |\n    41→| `-progress url` | Send machine-readable progress to url |\n    42→| `-report` | Dump full command line and log to a file |\n    43→| `-filter_threads n` | Number of threads for filter processing |\n    44→\n    45→### Input/Output Options\n    46→\n    47→| Flag | Description |\n    48→|------|-------------|\n    49→| `-i url` | Input file URL |\n    50→| `-f fmt` | Force input or output format |\n    51→| `-c[:stream] codec` | Select encoder/decoder; use `copy` for stream copying |\n    52→| `-t duration` | Limit duration (as input: read limit; as output: write limit) |\n    53→| `-to position` | Stop at position (timestamp) |\n    54→| `-ss position` | Seek to position (before `-i`: fast input seek; after: output seek) |\n    55→| `-sseof position` | Seek relative to end of file |\n    56→| `-itsoffset offset` | Set input time offset |\n    57→| `-itsscale scale` | Rescale input timestamps |\n    58→| `-metadata key=value` | Set metadata key/value pair |\n    59→| `-disposition value` | Set stream disposition flags |\n    60→| `-target type` | Specify target type: `vcd`, `svcd`, `dvd`, `dv`, `dv50` |\n    61→| `-stream_loop n` | Loop input stream n times (-1 = infinite) |\n    62→| `-frames[:stream] n` | Stop after n frames |\n    63→| `-fs limit` | Set file size limit in bytes |\n    64→| `-timestamp date` | Set recording timestamp |\n    65→\n    66→### Video Options\n    67→\n    68→| Flag | Description |\n    69→|------|-------------|\n    70→| `-vn` | Disable video |\n    71→| `-vcodec codec` | Set video codec (alias for `-c:v`) |\n    72→| `-r fps` | Set frame rate |\n    73→| `-fpsmax fps` | Set maximum frame rate |\n    74→| `-s WxH` | Set frame size |\n    75→| `-aspect ratio` | Set display aspect ratio (e.g. `16:9`) |\n    76→| `-pix_fmt format` | Set pixel format |\n    77→| `-vf filtergraph` | Apply video filter graph (alias for `-filter:v`) |\n    78→| `-pass n` | Two-pass encoding pass (1 or 2) |\n    79→| `-passlogfile prefix` | Two-pass log file prefix |\n    80→| `-vframes n` | Set number of video frames to output |\n    81→| `-autorotate` | Auto-rotate based on metadata (default on) |\n    82→| `-display_rotation angle` | Set video rotation metadata |\n    83→| `-display_hflip` | Horizontal flip metadata |\n    84→| `-display_vflip` | Vertical flip metadata |\n    85→| `-force_key_frames expr` | Force keyframes at specified times/expression |\n    86→| `-copyinkf` | Copy non-key frames at the beginning during stream copy |\n    87→\n    88→### Audio Options\n    89→\n    90→| Flag | Description |\n    91→|------|-------------|\n    92→| `-an` | Disable audio |\n    93→| `-acodec codec` | Set audio codec (alias for `-c:a`) |\n    94→| `-ar freq` | Set audio sample rate (Hz) |\n    95→| `-ac channels` | Set number of audio channels |\n    96→| `-af filtergraph` | Apply audio filter graph (alias for `-filter:a`) |\n    97→| `-sample_fmt fmt` | Set audio sample format |\n    98→| `-channel_layout layout` | Set audio channel layout |\n    99→| `-aq q` | Set audio quality (codec-specific VBR) |\n   100→| `-aframes n` | Set number of audio frames to output |\n   101→\n   102→### Subtitle Options\n   103→\n   104→| Flag | Description |\n   105→|------|-------------|\n   106→| `-sn` | Disable subtitles |\n   107→| `-scodec codec` | Set subtitle codec (alias for `-c:s`) |\n   108→| `-fix_sub_duration` | Fix subtitle durations to avoid overlap |\n   109→\n   110→### Stream Selection\n   111→\n   112→| Flag | Description |\n   113→|------|-------------|\n   114→| `-map input:stream` | Manually select streams for output |\n   115→| `-dn` | Disable data streams |\n   116→\n   117→Stream specifiers: `v` (video), `V` (video, no images), `a` (audio), `s` (subtitle), `d` (data). Index with `:N` (e.g. `a:0` = first audio).\n   118→\n   119→### Hardware Acceleration\n   120→\n   121→| Flag | Description |\n   122→|------|-------------|\n   123→| `-hwaccel method` | HW accel method: `cuda`, `vaapi`, `qsv`, `vulkan`, `auto` |\n   124→| `-hwaccel_device device` | Select HW device |\n   125→| `-init_hw_device type=name` | Initialize HW device |\n   126→\n   127→---\n   128→\n   129→## ffprobe\n   130→\n   131→### Synopsis\n   132→\n   133→```\n   134→ffprobe [options] input_url\n   135→```\n   136→\n   137→### Main Options\n   138→\n   139→| Flag | Description |\n   140→|------|-------------|\n   141→| `-show_format` | Show container format info |\n   142→| `-show_streams` | Show per-stream info |\n   143→| `-show_packets` | Show per-packet info |\n   144→| `-show_frames` | Show per-frame info |\n   145→| `-show_chapters` | Show chapter info |\n   146→| `-show_programs` | Show program info |\n   147→| `-show_entries section=key1,key2` | Show only specific fields |\n   148→| `-show_error` | Show probe errors |\n   149→| `-select_streams specifier` | Filter to specific streams (e.g. `v:0`, `a`) |\n   150→| `-count_frames` | Count frames per stream |\n   151→| `-count_packets` | Count packets per stream |\n   152→| `-read_intervals intervals` | Analyze specific time ranges |\n   153→\n   154→### Output Formats\n   155→\n   156→Set with `-output_format` (or `-of`, `-print_format`):\n   157→\n   158→| Format | Description |\n   159→|--------|-------------|\n   160→| `default` | `[SECTION] key=value [/SECTION]` |\n   161→| `json` | JSON output (most useful for parsing) |\n   162→| `xml` | XML output |\n   163→| `csv` | Comma-separated values |\n   164→| `flat` | Flat `key=value` per line |\n   165→| `ini` | INI-style sections |\n   166→\n   167→### Display Options\n   168→\n   169→| Flag | Description |\n   170→|------|-------------|\n   171→| `-pretty` | Human-readable units and time formatting |\n   172→| `-unit` | Show value units |\n   173→| `-sexagesimal` | Format times as HH:MM:SS.us |\n   174→| `-hide_banner` | Suppress copyright/build info |\n   175→| `-o output_url` | Write output to file instead of stdout |\n   176→\n   177→---\n   178→\n   179→## Common Codecs\n   180→\n   181→### Video Encoders\n   182→\n   183→#### libx264 (H.264)\n   184→\n   185→| Option | Description |\n   186→|--------|-------------|\n   187→| `-preset` | Speed/quality: `ultrafast`, `superfast`, `veryfast`, `faster`, `fast`, `medium` (default), `slow`, `slower`, `veryslow` |\n   188→| `-crf` | Constant quality: 0 (lossless) to 51 (worst). 18-23 is typical |\n   189→| `-profile:v` | `baseline`, `main`, `high` |\n   190→| `-tune` | `film`, `animation`, `grain`, `stillimage`, `fastdecode`, `zerolatency` |\n   191→| `-b:v` | Target bitrate (e.g. `2M`) |\n   192→\n   193→#### libx265 (H.265/HEVC)\n   194→\n   195→| Option | Description |\n   196→|--------|-------------|\n   197→| `-preset` | Same presets as x264 |\n   198→| `-crf` | 0-51, default 28. Similar quality to x264 at lower bitrate |\n   199→| `-profile:v` | `main`, `main10`, `main12` |\n   200→| `-b:v` | Target bitrate |\n   201→\n   202→#### libvpx-vp9 (VP9)\n   203→\n   204→| Option | Description |\n   205→|--------|-------------|\n   206→| `-crf` | 0-63. 31 is a good default |\n   207→| `-b:v` | Target bitrate (set to `0` for pure CRF mode) |\n   208→| `-cpu-used` | Speed: 0 (slowest/best) to 8 (fastest) |\n   209→| `-deadline` | `best`, `good` (default), `realtime` |\n   210→| `-row-mt 1` | Enable row-based multithreading |\n   211→\n   212→#### libsvtav1 (SVT-AV1)\n   213→\n   214→| Option | Description |\n   215→|--------|-------------|\n   216→| `-crf` | 0-63. 30 is a good default |\n   217→| `-preset` | 0 (slowest/best) to 13 (fastest). 8 is a good default |\n   218→| `-b:v` | Target bitrate |\n   219→\n   220→#### libaom-av1 (AOM AV1)\n   221→\n   222→| Option | Description |\n   223→|--------|-------------|\n   224→| `-crf` | 0-63 |\n   225→| `-cpu-used` | 0 (best) to 8 (fastest) |\n   226→| `-b:v` | Target bitrate (set to `0` for pure CRF mode) |\n   227→| `-tiles` | Tile columns x rows for parallelism |\n   228→\n   229→### Audio Encoders\n   230→\n   231→#### aac (Native AAC)\n   232→\n   233→| Option | Description |\n   234→|--------|-------------|\n   235→| `-b:a` | Bitrate: `128k`, `192k`, `256k` |\n   236→| `-profile:a` | `aac_low` (default), `aac_he`, `aac_he_v2` |\n   237→\n   238→#### libmp3lame (MP3)\n   239→\n   240→| Option | Description |\n   241→|--------|-------------|\n   242→| `-b:a` | CBR bitrate: `128k`, `192k`, `320k` |\n   243→| `-q:a` | VBR quality: 0 (best) to 9 (worst). 2 is a good default |\n   244→\n   245→#### libopus (Opus)\n   246→\n   247→| Option | Description |\n   248→|--------|-------------|\n   249→| `-b:a` | Bitrate: `64k` to `256k`. 128k is a good default |\n   250→| `-vbr` | `on` (default), `off`, `constrained` |\n   251→| `-application` | `audio` (default), `voip`, `lowdelay` |\n   252→\n   253→#### libvorbis (Vorbis)\n   254→\n   255→| Option | Description |\n   256→|--------|-------------|\n   257→| `-q:a` | VBR quality: -1 to 10. 5 is a good default |\n   258→| `-b:a` | ABR bitrate |\n   259→\n   260→#### flac (FLAC)\n   261→\n   262→| Option | Description |\n   263→|--------|-------------|\n   264→| `-compression_level` | 0 (fast) to 12 (best). 5 is default |\n   265→\n   266→---\n   267→\n   268→## Common Container Formats\n   269→\n   270→| Format | Extensions | Notes |\n   271→|--------|-----------|-------|\n   272→| `mp4` | .mp4, .m4a, .m4v | Use `-movflags +faststart` for web streaming |\n   273→| `matroska` | .mkv | Supports virtually all codecs |\n   274→| `webm` | .webm | VP8/VP9/AV1 + Vorbis/Opus for web |\n   275→| `avi` | .avi | Legacy; limited codec support |\n   276→| `mpegts` | .ts | Broadcast transport stream |\n   277→| `ogg` | .ogg, .ogv | Vorbis/Opus/Theora container |\n   278→| `wav` | .wav | Uncompressed PCM audio |\n   279→| `flac` | .flac | Lossless audio |\n   280→| `mp3` | .mp3 | MPEG audio layer 3 |\n   281→| `hls` | .m3u8 | HTTP Live Streaming |\n   282→| `dash` | .mpd | DASH adaptive streaming |\n   283→| `gif` | .gif | Animated GIF |\n   284→| `image2` | various | Image sequence input/output |\n   285→| `concat` | text file | Concatenation demuxer (file list) |\n   286→| `null` | — | Discard output (benchmarking) |\n   287→\n   288→---\n   289→\n   290→## Common Video Filters (`-vf`)\n   291→\n   292→| Filter | Description | Example |\n   293→|--------|-------------|---------|\n   294→| `scale=W:H` | Resize video. Use `-1` or `-2` to auto-calculate | `scale=1280:720`, `scale=-2:480` |\n   295→| `crop=W:H:X:Y` | Crop to WxH starting at X,Y | `crop=640:480:100:50` |\n   296→| `pad=W:H:X:Y:color` | Pad video with borders | `pad=1920:1080:(ow-iw)/2:(oh-ih)/2:black` |\n   297→| `overlay=X:Y` | Composite second input over first | `overlay=10:10` |\n   298→| `transpose=N` | Rotate: 0=90ccw+vflip, 1=90cw, 2=90ccw, 3=90cw+vflip | `transpose=1` |\n   299→| `hflip` / `vflip` | Horizontal / vertical flip | `hflip` |\n   300→| `rotate=angle` | Rotate by arbitrary angle (radians) | `rotate=PI/4` |\n   301→| `fps=N` | Change frame rate | `fps=30` |\n   302→| `setpts=expr` | Modify presentation timestamps | `setpts=0.5*PTS` (2x speed) |\n   303→| `trim=start:end` | Extract time range | `trim=start=10:end=20` |\n   304→| `drawtext=opts` | Overlay text | `drawtext=text='Hello':fontsize=24:x=10:y=10` |\n   305→| `fade=t=type:st=S:d=D` | Fade in/out | `fade=t=in:st=0:d=2` |\n   306→| `eq=opts` | Adjust brightness/contrast/saturation | `eq=brightness=0.1:contrast=1.2` |\n   307→| `format=pix_fmt` | Convert pixel format | `format=yuv420p` |\n   308→| `concat=n:v:a` | Concatenate segments | `concat=n=2:v=1:a=1` |\n   309→| `split` / `select` | Duplicate / select frames | `select='eq(pict_type,I)'` |\n   310→| `deinterlace` / `yadif` | Remove interlacing | `yadif=1` |\n   311→| `boxblur=R` | Apply box blur | `boxblur=5:1` |\n   312→| `subtitles=file` | Burn in subtitles from file | `subtitles=subs.srt` |\n   313→| `palettegen` / `paletteuse` | Generate/apply palette for GIF | Used in two-pass GIF creation |\n   314→| `colorchannelmixer` | Mix color channels | `colorchannelmixer=.3:.4:.3:0:.3:.4:.3:0:.3:.4:.3` (grayscale) |\n   315→\n   316→## Common Audio Filters (`-af`)\n   317→\n   318→| Filter | Description | Example |\n   319→|--------|-------------|---------|\n   320→| `volume=V` | Adjust volume | `volume=1.5`, `volume=-3dB` |\n   321→| `loudnorm` | EBU R128 loudness normalization | `loudnorm=I=-16:TP=-1.5:LRA=11` |\n   322→| `atempo=T` | Change tempo (0.5-100.0) | `atempo=2.0` (2x speed) |\n   323→| `aresample=rate` | Resample audio | `aresample=44100` |\n   324→| `amerge` | Merge audio channels | `amerge=inputs=2` |\n   325→| `afade=t=type:st=S:d=D` | Audio fade in/out | `afade=t=in:st=0:d=3` |\n   326→| `highpass=f=freq` | High-pass filter | `highpass=f=200` |\n   327→| `lowpass=f=freq` | Low-pass filter | `lowpass=f=3000` |\n   328→| `equalizer=f:t:w:g` | Parametric EQ | `equalizer=f=1000:t=q:w=1:g=5` |\n   329→| `acompressor` | Dynamic range compression | `acompressor=threshold=-20dB:ratio=4` |\n   330→| `silenceremove` | Remove silence | `silenceremove=1:0:-50dB` |\n   331→| `silencedetect` | Detect silence | `silencedetect=n=-30dB:d=2` |\n   332→| `adelay=delays` | Delay audio channels | `adelay=1000\\|1000` (ms) |\n   333→| `aecho=id:ig:delays:decays` | Add echo effect | `aecho=0.8:0.88:60:0.4` |\n   334→| `pan=layout:gains` | Remix channels | `pan=mono\\|c0=0.5*c0+0.5*c1` |\n   335→\n   336→---\n   337→\n   338→## Common Recipes\n   339→\n   340→### Convert format\n   341→\n   342→```bash\n   343→ffmpeg -i input.mkv output.mp4\n   344→```\n   345→\n   346→### Transcode with CRF quality\n   347→\n   348→```bash\n   349→ffmpeg -i input.mp4 -c:v libx264 -crf 20 -c:a aac -b:a 192k output.mp4\n   350→```\n   351→\n   352→### Extract audio\n   353→\n   354→```bash\n   355→ffmpeg -i video.mp4 -vn -c:a copy audio.m4a\n   356→```\n   357→\n   358→### Extract video (no audio)\n   359→\n   360→```bash\n   361→ffmpeg -i input.mp4 -an -c:v copy output.mp4\n   362→```\n   363→\n   364→### Trim / cut\n   365→\n   366→```bash\n   367→ffmpeg -ss 00:01:30 -to 00:03:00 -i input.mp4 -c copy output.mp4\n   368→```\n   369→\n   370→### Resize video\n   371→\n   372→```bash\n   373→ffmpeg -i input.mp4 -vf \"scale=1280:720\" -c:a copy output.mp4\n   374→```\n   375→\n   376→### Compress video (lower quality)\n   377→\n   378→```bash\n   379→ffmpeg -i input.mp4 -c:v libx264 -crf 28 -preset fast -c:a aac -b:a 128k output.mp4\n   380→```\n   381→\n   382→### Two-pass encoding\n   383→\n   384→```bash\n   385→ffmpeg -i input.mp4 -c:v libx264 -b:v 2M -pass 1 -f null /dev/null\n   386→ffmpeg -i input.mp4 -c:v libx264 -b:v 2M -pass 2 output.mp4\n   387→```\n   388→\n   389→### Concatenate files (concat demuxer)\n   390→\n   391→```bash\n   392→# files.txt contains:\n   393→# file 'part1.mp4'\n   394→# file 'part2.mp4'\n   395→ffmpeg -f concat -safe 0 -i files.txt -c copy output.mp4\n   396→```\n   397→\n   398→### Add subtitles (burn-in)\n   399→\n   400→```bash\n   401→ffmpeg -i input.mp4 -vf \"subtitles=subs.srt\" output.mp4\n   402→```\n   403→\n   404→### Create GIF\n   405→\n   406→```bash\n   407→ffmpeg -i input.mp4 -vf \"fps=10,scale=320:-1:flags=lanczos,split[s0][s1];[s0]palettegen[p];[s1][p]paletteuse\" output.gif\n   408→```\n   409→\n   410→### Add watermark / overlay\n   411→\n   412→```bash\n   413→ffmpeg -i video.mp4 -i logo.png -filter_complex \"overlay=10:10\" output.mp4\n   414→```\n   415→\n   416→### Change speed (video + audio)\n   417→\n   418→```bash\n   419→ffmpeg -i input.mp4 -vf \"setpts=0.5*PTS\" -af \"atempo=2.0\" output.mp4\n   420→```\n   421→\n   422→### Extract frames as images\n   423→\n   424→```bash\n   425→ffmpeg -i input.mp4 -vf \"fps=1\" frame_%04d.png\n   426→```\n   427→\n   428→### Merge audio and video\n   429→\n   430→```bash\n   431→ffmpeg -i video.mp4 -i audio.m4a -c:v copy -c:a copy -shortest output.mp4\n   432→```\n   433→\n   434→### Normalize audio loudness\n   435→\n   436→```bash\n   437→ffmpeg -i input.mp4 -af \"loudnorm=I=-16:TP=-1.5:LRA=11\" -c:v copy output.mp4\n   438→```\n   439→\n   440→### Convert to web-optimized MP4\n   441→\n   442→```bash\n   443→ffmpeg -i input.mp4 -c:v libx264 -crf 23 -preset medium -c:a aac -b:a 128k -movflags +faststart output.mp4\n   444→```\n   445→\n   446→### Probe file info (JSON)\n   447→\n   448→```bash\n   449→ffprobe -v quiet -print_format json -show_format -show_streams input.mp4\n   450→```\n   451→\n   452→### Get duration only\n   453→\n   454→```bash\n   455→ffprobe -v quiet -show_entries format=duration -of default=noprint_wrappers=1:nokey=1 input.mp4\n   456→```\n   457→\n   458→### Get resolution only\n   459→\n   460→```bash\n   461→ffprobe -v quiet -select_streams v:0 -show_entries stream=width,height -of csv=s=x:p=0 input.mp4\n   462→```\n   463→\n   464→---\n   465→\n   466→## Source\n   467→\n   468→- Website: https://ffmpeg.org/\n   469→- Documentation: https://www.ffmpeg.org/documentation.html\n   470→- CLI reference: https://www.ffmpeg.org/ffmpeg.html\n   471→- Filters reference: https://www.ffmpeg.org/ffmpeg-filters.html\n   472→- Codecs reference: https://www.ffmpeg.org/ffmpeg-codecs.html\n   473→- Formats reference: https://www.ffmpeg.org/ffmpeg-formats.html\n   474→- Wiki: https://trac.ffmpeg.org/wiki\n   475→","isError":false},{"role":"tool","toolName":"WebFetch","toolInput":{"url":"https://github.com/Blaizzy/mlx-audio","prompt":"Extract all information about mlx-audio: what it is, installation instructions, CLI usage, Python API usage, all available commands, supported models, supported features (TTS, STT, voice activity detection, etc.), configuration options, and any examples. Be comprehensive."},"toolUseId":"toolu_01BTviqNNX8vHWJ2Ci1yB8xS","output":"# MLX-Audio: Comprehensive Overview\n\n## What It Is\n\nMLX-Audio is a speech processing library built on Apple's MLX framework. It provides \"fast and efficient text-to-speech (TTS), speech-to-text (STT), and speech-to-speech (STS)\" capabilities optimized for Apple Silicon devices.\n\n## Installation\n\n**Via pip:**\n```bash\npip install mlx-audio\n```\n\n**CLI tools with uv (latest from PyPI):**\n```bash\nuv tool install --force mlx-audio --prerelease=allow\n```\n\n**Development setup:**\n```bash\ngit clone https://github.com/Blaizzy/mlx-audio.git\ncd mlx-audio\npip install -e \".[dev]\"\n```\n\n## Key Features\n\nThe library offers \"fast inference optimized for Apple Silicon (M series chips)\" along with multilingual support, voice customization, quantization options (3-bit through 8-bit), and an OpenAI-compatible REST API.\n\n## CLI Commands\n\n### TTS Generation\n```bash\nmlx_audio.tts.generate --model mlx-community/Kokoro-82M-bf16 --text 'Hello, world!' --lang_code a\nmlx_audio.tts.generate --model mlx-community/Kokoro-82M-bf16 --text 'Hello!' --voice af_heart --speed 1.2 --lang_code a\nmlx_audio.tts.generate --model mlx-community/Kokoro-82M-bf16 --text 'Hello!' --play --lang_code a\nmlx_audio.tts.generate --model mlx-community/Kokoro-82M-bf16 --text 'Hello!' --output_path ./my_audio --lang_code a\n```\n\n### STT Transcription\n```bash\npython -m mlx_audio.stt.generate --model mlx-community/VibeVoice-ASR-bf16 --audio meeting.wav --output-path output --format json --max-tokens 8192 --verbose\npython -m mlx_audio.stt.generate --model mlx-community/parakeet-tdt-0.6b-v3 --audio speech.wav --output-path output --format json --verbose\n```\n\n## Python API Usage\n\n### TTS Example\n```python\nfrom mlx_audio.tts.utils import load_model\n\nmodel = load_model(\"mlx-community/Kokoro-82M-bf16\")\nfor result in model.generate(\"Hello from MLX-Audio!\", voice=\"af_heart\"):\n    audio = result.audio  # mx.array waveform\n```\n\n### STT Example\n```python\nfrom mlx_audio.stt.generate import generate_transcription\n\nresult = generate_transcription(\n    model=\"mlx-community/whisper-large-v3-turbo-asr-fp16\",\n    audio=\"audio.wav\",\n)\nprint(result.text)\n```\n\n### Qwen3-ASR with Forced Alignment\n```python\nfrom mlx_audio.stt import load\n\nmodel = load(\"mlx-community/Qwen3-ASR-0.6B-8bit\")\nresult = model.generate(\"audio.wav\", language=\"English\")\nprint(result.text)\n\naligner = load(\"mlx-community/Qwen3-ForcedAligner-0.6B-8bit\")\nresult = aligner.generate(\"audio.wav\", text=\"I have a dream\", language=\"English\")\nfor item in result:\n    print(f\"[{item.start_time:.2f}s - {item.end_time:.2f}s] {item.text}\")\n```\n\n### Voice Cloning (CSM)\n```bash\nmlx_audio.tts.generate --model mlx-community/csm-1b --text \"Hello from Sesame.\" --ref_audio ./reference_voice.wav --play\n```\n\n### VibeVoice-ASR with Context\n```python\nresult = model.generate(\n    audio=\"technical_talk.wav\",\n    context=\"MLX, Apple Silicon, PyTorch, Transformer\",\n    max_tokens=8192,\n    temperature=0.0,\n)\n```\n\n### Streaming Transcription\n```python\n# VibeVoice-ASR streaming\nfor text in model.stream_transcribe(audio=\"speech.wav\", max_tokens=4096):\n    print(text, end=\"\", flush=True)\n\n# Parakeet streaming\nfor chunk in model.generate(\"long_audio.wav\", stream=True):\n    print(chunk.text, end=\"\", flush=True)\n\n# Voxtral Realtime streaming\nfor chunk in model.generate(\"audio.wav\", stream=True):\n    print(chunk, end=\"\", flush=True)\n```\n\n## Supported TTS Models\n\n| Model | Description | Languages |\n|-------|-------------|-----------|\n| **Kokoro** | Fast, high-quality multilingual | EN, JA, ZH, FR, ES, IT, PT, HI |\n| **Qwen3-TTS** | Alibaba's multilingual with voice design | ZH, EN, JA, KO, + more |\n| **CSM** | Voice cloning | EN |\n| **Dia** | Dialogue-focused | EN |\n| **OuteTTS** | Efficient model | EN |\n| **Spark** | SparkTTS | EN, ZH |\n| **Chatterbox** | Expressive multilingual | EN, ES, FR, DE, IT, PT, PL, TR, RU, NL, CS, AR, ZH, JA, HU, KO |\n| **Soprano** | High-quality | EN |\n\n**Kokoro Voice Examples:** af_heart, af_bella, af_nova, af_sky (American female); am_adam, am_echo (American male); bf_alice, bf_emma (British female); bm_daniel, bm_george (British male); jf_alpha, jm_kumo (Japanese); zf_xiaobei, zm_yunxi (Chinese)\n\n**Language Codes:** a (American English), b (British English), j (Japanese), z (Mandarin Chinese), e (Spanish), f (French)\n\n## Supported STT Models\n\n| Model | Description | Languages |\n|-------|-------------|-----------|\n| **Whisper** | OpenAI's robust model | 99+ languages |\n| **Qwen3-ASR** | Alibaba's multilingual | ZH, EN, JA, KO, + more |\n| **Qwen3-ForcedAligner** | Word-level alignment | ZH, EN, JA, KO, + more |\n| **Parakeet** | NVIDIA's accurate STT | EN (v2), 25 EU languages (v3) |\n| **Voxtral** | Mistral's speech model | Multiple |\n| **Voxtral Realtime** | 4B streaming STT | Multiple |\n| **VibeVoice-ASR** | Microsoft's 9B with diarization | Multiple |\n\n**Parakeet v3 Languages:** Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Slovak, Slovenian, Spanish, Swedish, Russian, Ukrainian\n\n## VAD/Speaker Diarization Models\n\n| Model | Description | \n|-------|-------------|\n| **Sortformer v1** | NVIDIA's end-to-end (up to 4 speakers) |\n| **Sortformer v2.1** | Streaming diarization with AOSC compression |\n\n## Speech-to-Speech (STS) Models\n\n| Model | Purpose |\n|-------|---------|\n| **SAM-Audio** | Text-guided source separation |\n| **Liquid2.5-Audio** | Speech/text-to-speech and STT |\n| **MossFormer2 SE** | Speech enhancement/noise removal |\n\n## Configuration Options\n\n### TTS Parameters\n- `--text`: Input text to synthesize\n- `--voice`: Voice preset selection\n- `--speed`: Speech speed adjustment (e.g., 1.2)\n- `--lang_code`: Language code\n- `--play`: Play audio immediately\n- `--output_path`: Directory for saving audio\n- `--ref_audio`: Reference audio for voice cloning\n\n### STT Parameters\n- `--audio`: Input audio file\n- `--model`: Model identifier\n- `--language`: Language specification\n- `--max_tokens`: Token limit (e.g., 8192)\n- `--temperature`: Generation temperature (0.0 for deterministic)\n- `--context`: Hotwords/metadata for context\n- `--output_path`: Output directory\n- `--format`: Output format (json, etc.)\n- `--stream`: Enable streaming mode\n- `--transcription_delay_ms`: Latency setting (Voxtral Realtime)\n- `--verbose`: Detailed logging\n\n## License\n\nThe project is licensed under MIT.","isError":false},{"role":"assistant","text":"Now let me fetch the README more thoroughly, particularly for details about the API server and quantization."},{"role":"tool","toolName":"WebFetch","toolInput":{"url":"https://github.com/Blaizzy/mlx-audio/blob/main/README.md","prompt":"Extract the complete raw content of the README. Focus on: 1) The REST API / OpenAI-compatible server setup and usage, 2) Quantization details, 3) Any additional CLI flags or Python API details not covered before. Get everything."},"toolUseId":"toolu_019imtWTE2Lakb4ZaZMFMcD9","output":"# MLX-Audio README: REST API, Quantization & Advanced Features\n\n## Web Interface & API Server\n\nThe project supports an OpenAI-compatible REST API server. Starting the server enables both a web interface with 3D audio visualization and programmatic access to all audio processing models.\n\n**Starting the Server:**\nLaunch the API server with a simple command to make all TTS, STT, and other audio models accessible via HTTP endpoints.\n\n**API Endpoints:**\nThe server implements OpenAI-compatible REST endpoints, allowing developers to integrate MLX-Audio into existing applications designed for OpenAI's API. This includes standard endpoints for speech synthesis, transcription, and other audio operations.\n\n## Quantization\n\nThe library supports multiple quantization levels: 3-bit, 4-bit, 6-bit, 8-bit, and additional precision options. Quantization reduces model size and improves inference speed on Apple Silicon while maintaining quality. Users can select pre-quantized model variants from Hugging Face (e.g., \"mlx-community/model-name-4bit\") or quantize models locally.\n\n## Additional CLI & Python API Features\n\n- **Speech speed control**: Adjust playback speed for TTS generation\n- **Output path specification**: Save generated audio to custom directories\n- **Immediate playback**: Use `--play` flag to hear results instantly\n- **Streaming support**: Several models support streaming transcription for real-time processing\n- **Context/hotwords**: VibeVoice-ASR and other models accept context strings to improve technical term recognition\n- **Model variants**: Choose between different precision levels and model sizes for performance/quality tradeoffs\n\nThe library emphasizes ease of use across command-line, Python API, and web interface paradigms for audio processing workflows.","isError":false},{"role":"tool","toolName":"WebFetch","toolInput":{"url":"https://github.com/Blaizzy/mlx-audio/tree/main/mlx_audio/tts","prompt":"List all files and directories in the TTS module. Look for any server.py, api.py, or similar files that indicate REST API functionality."},"toolUseId":"toolu_01VNu3dPPG68dArCP19kaXH9","output":"# TTS Module Contents\n\nBased on the GitHub page content, the TTS (Text-to-Speech) module directory structure includes:\n\n**Files:**\n- `__init__.py`\n- `audio_player.py`\n- `generate.py`\n- `utils.py`\n\n**Subdirectories:**\n- `models/`\n- `tests/`\n\n## API Functionality Assessment\n\nThere is **no evidence of REST API files** (such as `server.py` or `api.py`) within the TTS module itself. The module appears focused on core TTS functionality through generation, audio playback, and utility functions rather than exposing HTTP endpoints.\n\n**Note:** The parent `mlx_audio` directory does contain a `server.py` file, which may provide API capabilities at the package level, but this is separate from the TTS module's direct implementation.","isError":false},{"role":"tool","toolName":"WebFetch","toolInput":{"url":"https://github.com/Blaizzy/mlx-audio/blob/main/mlx_audio/server.py","prompt":"Extract the full content of this server.py file. I need to understand: 1) How to start the server (command), 2) All available API endpoints and their parameters, 3) Request/response formats, 4) Any configuration options. Get everything."},"toolUseId":"toolu_017WJRxSMxEMSVeMGNLcgC22","output":"# MLX Audio Server - Complete Documentation\n\n## Starting the Server\n\n**Command:**\n```bash\npython -m mlx_audio.server [OPTIONS]\n```\n\n**Key startup options:**\n- `--host`: Server host (default: localhost)\n- `--port`: Server port (default: 8000)\n- `--allowed-origins`: CORS allowed origins (default: *)\n- `--workers`: Number of workers (int or float, default: 2)\n- `--reload`: Enable auto-reload\n- `--start-ui`: Launch Studio UI alongside API\n- `--log-dir`: Directory for server logs (default: logs)\n\n## API Endpoints\n\n### Model Management\n\n**GET `/v1/models`**\n- Returns list of available models in OpenAI-compatible format\n- Response: `{\"object\": \"list\", \"data\": [{\"id\": model_name, \"object\": \"model\", ...}]}`\n\n**POST `/v1/models`**\n- Adds new model to API\n- Parameter: `model_name` (query string)\n- Response: `{\"status\": \"success\", \"message\": \"Model {name} added successfully\"}`\n\n**DELETE `/v1/models`**\n- Removes model from API\n- Parameter: `model_name` (query string)\n- Response: 204 No Content on success, 404 if not found\n\n### Speech Generation (TTS)\n\n**POST `/v1/audio/speech`**\n- Generates speech from text\n- Request body (SpeechRequest):\n```\n{\n  \"model\": str (required),\n  \"input\": str (required),\n  \"voice\": str | None,\n  \"speed\": float (default: 1.0),\n  \"gender\": str (default: \"male\"),\n  \"pitch\": float (default: 1.0),\n  \"instruct\": str | None,\n  \"lang_code\": str (default: \"a\"),\n  \"ref_audio\": str | None,\n  \"ref_text\": str | None,\n  \"temperature\": float (default: 0.7),\n  \"top_p\": float (default: 0.95),\n  \"top_k\": int (default: 40),\n  \"repetition_penalty\": float (default: 1.0),\n  \"response_format\": str (default: \"mp3\"),\n  \"stream\": bool (default: False),\n  \"streaming_interval\": float (default: 2.0),\n  \"max_tokens\": int (default: 1200),\n  \"verbose\": bool (default: False)\n}\n```\n- Response: Audio stream with appropriate content-type\n\n### Speech-to-Text (STT)\n\n**POST `/v1/audio/transcriptions`**\n- Transcribes audio file\n- Parameters (form-data):\n  - `file`: Audio file (required)\n  - `model`: Model name (required)\n  - `language`: Optional language code\n  - `verbose`: Boolean (default: False)\n  - `max_tokens`: Integer (default: 1024)\n  - `chunk_duration`: Float (default: 30.0)\n  - `frame_threshold`: Integer (default: 25)\n  - `stream`: Boolean (default: False)\n  - `context`: Optional context string\n  - `prefill_step_size`: Integer (default: 2048)\n  - `text`: Optional text parameter\n\n- Response: NDJSON stream with transcription chunks\n```\n{\"text\": str, \"accumulated\": str}\nor\n{\"text\": str, \"start\": float, \"end\": float, \"is_final\": bool, \"language\": str}\n```\n\n**WebSocket `/v1/audio/transcriptions/realtime`**\n- Real-time transcription via WebSocket\n- Initial message (JSON):\n```\n{\n  \"model\": str,\n  \"language\": str | None,\n  \"sample_rate\": int (default: 16000),\n  \"streaming\": bool (default: True)\n}\n```\n- Audio: Send as binary frames (int16 PCM)\n- Responses:\n  - `{\"status\": \"ready\", \"message\": \"...\"}`\n  - `{\"type\": \"delta\", \"delta\": str}`\n  - `{\"type\": \"complete\", \"text\": str, \"segments\": null, ...}`\n\n### Audio Separation\n\n**POST `/v1/audio/separations`**\n- Separates audio using SAM Audio model\n- Parameters (form-data):\n  - `file`: Audio file (required)\n  - `model`: Model name (default: mlx-community/sam-audio-large-fp16)\n  - `description`: Target description (default: \"speech\")\n  - `method`: ODE method \"midpoint\" or \"euler\" (default: midpoint)\n  - `steps`: ODE steps 2/4/8/16/32 (default: 16)\n\n- Response (SeparationResponse):\n```\n{\n  \"target\": \"base64_encoded_wav\",\n  \"residual\": \"base64_encoded_wav\",\n  \"sample_rate\": int\n}\n```\n\n### Root\n\n**GET `/`**\n- Welcome message with link to UI (http://localhost:3000)\n\n## Configuration\n\n**Environment Variables:**\n- `MLX_AUDIO_NUM_WORKERS`: Override worker count\n- `MLX_AUDIO_ALLOWED_ORIGINS`: Comma-separated CORS origins\n\n**CORS Middleware:**\nConfigured automatically on startup; can be reconfigured via `setup_cors(app, origins_list)`\n\n## Architecture Notes\n\n- Uses FastAPI with async/await\n- Implements asyncio locks for thread-safe model management\n- VAD (Voice Activity Detection) enabled for real-time transcription\n- Supports streaming and non-streaming inference modes\n- Sanitizes NaN/Infinity values for JSON compatibility","isError":false},{"role":"tool","toolName":"WebFetch","toolInput":{"url":"https://github.com/Blaizzy/mlx-audio/tree/main/mlx_audio","prompt":"List all files and directories in the mlx_audio root package directory."},"toolUseId":"toolu_01TfedYLiRZqsWufnn4awqmw","output":"# Files and Directories in mlx_audio Root Package\n\nBased on the repository structure provided, the mlx_audio root package contains:\n\n**Directories:**\n- codec\n- sts\n- stt\n- tests\n- tts\n- ui\n- vad\n\n**Files:**\n- `__init__.py`\n- `audio_io.py`\n- `base.py`\n- `convert.py`\n- `dsp.py`\n- `server.py`\n- `utils.py`\n- `version.py`\n\nThis appears to be an audio processing library with modules for speech-to-text (stt), text-to-speech (tts), voice activity detection (vad), speech translation (sts), and codec handling.","isError":false},{"role":"assistant","text":"Now I have comprehensive information. Let me write the SKILL.md file."},{"role":"tool","toolName":"Write","toolInput":{"file_path":"/home/pastilhas/projects/officer.dev/data/skills/mlxaudio/SKILL.md","content":"---\nname: mlx.audio\ndescription: Generate speech from text and transcribe audio using mlx-audio. Use when the user wants text-to-speech synthesis, speech-to-text transcription, voice cloning, audio separation, or speech-to-speech processing on Apple Silicon.\n---\n\n# MLX-Audio\n\nA speech processing library built on Apple's MLX framework, providing TTS, STT, speech-to-speech (STS), and audio separation optimized for Apple Silicon.\n\n- **Repository:** https://github.com/Blaizzy/mlx-audio\n- **License:** MIT\n\n## CLI Tools\n\n### Text-to-Speech (TTS)\n\n```bash\nmlx_audio.tts.generate --model <model> --text '<text>' [options]\n```\n\n| Flag | Type | Default | Description |\n|------|------|---------|-------------|\n| `--model` | string | required | HuggingFace model ID |\n| `--text` | string | required | Text to synthesize |\n| `--voice` | string | — | Voice preset (model-specific) |\n| `--speed` | float | 1.0 | Speech speed multiplier |\n| `--lang_code` | string | `a` | Language code |\n| `--play` | flag | — | Play audio immediately |\n| `--output_path` | string | — | Directory to save audio |\n| `--ref_audio` | string | — | Reference audio for voice cloning (CSM) |\n\n#### Language Codes\n\n| Code | Language |\n|------|----------|\n| `a` | American English |\n| `b` | British English |\n| `j` | Japanese |\n| `z` | Mandarin Chinese |\n| `e` | Spanish |\n| `f` | French |\n\n#### Kokoro Voices\n\n| Voice | Description |\n|-------|-------------|\n| `af_heart`, `af_bella`, `af_nova`, `af_sky` | American female |\n| `am_adam`, `am_echo` | American male |\n| `bf_alice`, `bf_emma` | British female |\n| `bm_daniel`, `bm_george` | British male |\n| `jf_alpha`, `jm_kumo` | Japanese |\n| `zf_xiaobei`, `zm_yunxi` | Chinese |\n\n#### Examples\n\n```bash\n# Basic generation\nmlx_audio.tts.generate --model mlx-community/Kokoro-82M-bf16 --text 'Hello, world!' --lang_code a\n\n# With voice and speed\nmlx_audio.tts.generate --model mlx-community/Kokoro-82M-bf16 --text 'Hello!' --voice af_heart --speed 1.2 --lang_code a\n\n# Play immediately\nmlx_audio.tts.generate --model mlx-community/Kokoro-82M-bf16 --text 'Hello!' --play --lang_code a\n\n# Voice cloning with CSM\nmlx_audio.tts.generate --model mlx-community/csm-1b --text \"Hello from Sesame.\" --ref_audio ./reference_voice.wav --play\n```\n\n### Speech-to-Text (STT)\n\n```bash\npython -m mlx_audio.stt.generate --model <model> --audio <file> [options]\n```\n\n| Flag | Type | Default | Description |\n|------|------|---------|-------------|\n| `--model` | string | required | HuggingFace model ID |\n| `--audio` | string | required | Input audio file |\n| `--language` | string | — | Language code |\n| `--max-tokens` | int | 1024 | Maximum output tokens |\n| `--temperature` | float | 0.0 | Sampling temperature |\n| `--context` | string | — | Hotwords/metadata for context |\n| `--output-path` | string | — | Output directory |\n| `--format` | string | — | Output format (e.g. `json`) |\n| `--stream` | flag | — | Enable streaming mode |\n| `--verbose` | flag | — | Detailed logging |\n\n#### Examples\n\n```bash\n# Basic transcription\npython -m mlx_audio.stt.generate --model mlx-community/whisper-large-v3-turbo-asr-fp16 --audio speech.wav --verbose\n\n# With context for technical terms\npython -m mlx_audio.stt.generate --model mlx-community/VibeVoice-ASR-bf16 --audio meeting.wav --context \"MLX, Apple Silicon, PyTorch\" --max-tokens 8192 --format json --verbose\n\n# Parakeet model\npython -m mlx_audio.stt.generate --model mlx-community/parakeet-tdt-0.6b-v3 --audio speech.wav --format json --verbose\n```\n\n---\n\n## Python API\n\n### TTS\n\n```python\nfrom mlx_audio.tts.utils import load_model\n\nmodel = load_model(\"mlx-community/Kokoro-82M-bf16\")\nfor result in model.generate(\"Hello from MLX-Audio!\", voice=\"af_heart\"):\n    audio = result.audio  # mx.array waveform\n```\n\n### STT\n\n```python\nfrom mlx_audio.stt.generate import generate_transcription\n\nresult = generate_transcription(\n    model=\"mlx-community/whisper-large-v3-turbo-asr-fp16\",\n    audio=\"audio.wav\",\n)\nprint(result.text)\n```\n\n### STT with Streaming\n\n```python\nfrom mlx_audio.stt import load\n\n# VibeVoice-ASR streaming\nmodel = load(\"mlx-community/VibeVoice-ASR-bf16\")\nfor text in model.stream_transcribe(audio=\"speech.wav\", max_tokens=4096):\n    print(text, end=\"\", flush=True)\n\n# Parakeet streaming\nmodel = load(\"mlx-community/parakeet-tdt-0.6b-v3\")\nfor chunk in model.generate(\"long_audio.wav\", stream=True):\n    print(chunk.text, end=\"\", flush=True)\n```\n\n### Forced Alignment (Qwen3)\n\n```python\nfrom mlx_audio.stt import load\n\naligner = load(\"mlx-community/Qwen3-ForcedAligner-0.6B-8bit\")\nresult = aligner.generate(\"audio.wav\", text=\"I have a dream\", language=\"English\")\nfor item in result:\n    print(f\"[{item.start_time:.2f}s - {item.end_time:.2f}s] {item.text}\")\n```\n\n---\n\n## REST API Server (OpenAI-compatible)\n\n### Starting the Server\n\n```bash\npython -m mlx_audio.server [OPTIONS]\n```\n\n| Flag | Type | Default | Description |\n|------|------|---------|-------------|\n| `--host` | string | `localhost` | Server host |\n| `--port` | int | `8000` | Server port |\n| `--allowed-origins` | string | `*` | CORS allowed origins |\n| `--workers` | int/float | `2` | Number of workers |\n| `--reload` | flag | — | Enable auto-reload |\n| `--start-ui` | flag | — | Launch Studio UI alongside API |\n| `--log-dir` | string | `logs` | Directory for server logs |\n\n### Endpoints\n\n#### GET /v1/models\n\nList available models.\n\n```bash\ncurl http://localhost:8000/v1/models\n```\n\n#### POST /v1/models?model_name=\\<name\\>\n\nAdd a model to the server.\n\n#### DELETE /v1/models?model_name=\\<name\\>\n\nRemove a model from the server.\n\n#### POST /v1/audio/speech\n\nGenerate speech from text.\n\n```bash\ncurl -X POST http://localhost:8000/v1/audio/speech \\\n  -H \"Content-Type: application/json\" \\\n  -d '{\n    \"model\": \"mlx-community/Kokoro-82M-bf16\",\n    \"input\": \"Hello, world!\",\n    \"voice\": \"af_heart\",\n    \"speed\": 1.0,\n    \"lang_code\": \"a\",\n    \"response_format\": \"mp3\"\n  }' --output speech.mp3\n```\n\n| Parameter | Type | Default | Description |\n|-----------|------|---------|-------------|\n| `model` | string | required | Model ID |\n| `input` | string | required | Text to synthesize |\n| `voice` | string | — | Voice preset |\n| `speed` | float | 1.0 | Speech speed |\n| `lang_code` | string | `a` | Language code |\n| `ref_audio` | string | — | Reference audio path (voice cloning) |\n| `ref_text` | string | — | Reference transcript |\n| `response_format` | string | `mp3` | Output format |\n| `stream` | bool | false | Enable streaming |\n| `streaming_interval` | float | 2.0 | Streaming chunk interval |\n| `temperature` | float | 0.7 | Sampling temperature |\n| `top_p` | float | 0.95 | Nucleus sampling |\n| `top_k` | int | 40 | Top-k sampling |\n| `repetition_penalty` | float | 1.0 | Repetition penalty |\n| `max_tokens` | int | 1200 | Maximum tokens |\n| `gender` | string | `male` | Gender hint |\n| `pitch` | float | 1.0 | Pitch adjustment |\n| `instruct` | string | — | Instruction text |\n\n#### POST /v1/audio/transcriptions\n\nTranscribe an audio file (multipart/form-data).\n\n```bash\ncurl -X POST http://localhost:8000/v1/audio/transcriptions \\\n  -F file=@audio.wav \\\n  -F model=mlx-community/whisper-large-v3-turbo-asr-fp16\n```\n\n| Parameter | Type | Default | Description |\n|-----------|------|---------|-------------|\n| `file` | file | required | Audio file |\n| `model` | string | required | Model ID |\n| `language` | string | — | Language code |\n| `max_tokens` | int | 1024 | Maximum tokens |\n| `chunk_duration` | float | 30.0 | Chunk duration (seconds) |\n| `stream` | bool | false | Enable streaming |\n| `context` | string | — | Hotwords/context |\n| `text` | string | — | Reference text |\n| `verbose` | bool | false | Detailed output |\n\nResponse (NDJSON stream):\n\n```json\n{\"text\": \"chunk text\", \"accumulated\": \"full text so far\"}\n```\n\n#### WebSocket /v1/audio/transcriptions/realtime\n\nReal-time transcription via WebSocket. Send initial config as JSON, then stream int16 PCM audio as binary frames.\n\n```json\n{\n  \"model\": \"mlx-community/whisper-large-v3-turbo-asr-fp16\",\n  \"sample_rate\": 16000,\n  \"streaming\": true\n}\n```\n\n#### POST /v1/audio/separations\n\nSeparate audio sources (multipart/form-data).\n\n```bash\ncurl -X POST http://localhost:8000/v1/audio/separations \\\n  -F file=@audio.wav \\\n  -F model=mlx-community/sam-audio-large-fp16 \\\n  -F description=\"speech\"\n```\n\n| Parameter | Type | Default | Description |\n|-----------|------|---------|-------------|\n| `file` | file | required | Audio file |\n| `model` | string | `mlx-community/sam-audio-large-fp16` | Model ID |\n| `description` | string | `speech` | Target description |\n| `method` | string | `midpoint` | ODE method (`midpoint` or `euler`) |\n| `steps` | int | 16 | ODE steps (2/4/8/16/32) |\n\nResponse:\n\n```json\n{\n  \"target\": \"<base64 WAV>\",\n  \"residual\": \"<base64 WAV>\",\n  \"sample_rate\": 44100\n}\n```\n\n---\n\n## Supported Models\n\n### TTS Models\n\n| Model | Languages | Notes |\n|-------|-----------|-------|\n| Kokoro | EN, JA, ZH, FR, ES, IT, PT, HI | Fast, high-quality multilingual |\n| Qwen3-TTS | ZH, EN, JA, KO, + more | Voice design via instruction |\n| CSM | EN | Voice cloning with reference audio |\n| Dia | EN | Dialogue-focused |\n| OuteTTS | EN | Efficient |\n| Spark | EN, ZH | SparkTTS |\n| Chatterbox | EN, ES, FR, DE, IT, PT, PL, TR, RU, NL, CS, AR, ZH, JA, HU, KO | Expressive multilingual |\n| Soprano | EN | High-quality |\n\n### STT Models\n\n| Model | Languages | Notes |\n|-------|-----------|-------|\n| Whisper | 99+ languages | OpenAI's robust model |\n| Qwen3-ASR | ZH, EN, JA, KO, + more | Alibaba multilingual |\n| Qwen3-ForcedAligner | ZH, EN, JA, KO, + more | Word-level alignment |\n| Parakeet | EN (v2), 25 EU languages (v3) | NVIDIA, high accuracy |\n| Voxtral | Multiple | Mistral speech model |\n| Voxtral Realtime | Multiple | 4B streaming STT |\n| VibeVoice-ASR | Multiple | Microsoft 9B, supports diarization and context |\n\n### Other Models\n\n| Model | Type | Description |\n|-------|------|-------------|\n| Sortformer v1/v2.1 | VAD/Diarization | Speaker diarization (up to 4 speakers) |\n| SAM-Audio | Separation | Text-guided source separation |\n| Liquid2.5-Audio | STS | Speech/text-to-speech and STT |\n| MossFormer2 SE | Enhancement | Speech enhancement / noise removal |\n\nModels are available from `mlx-community` on HuggingFace with various quantization levels (3-bit through 8-bit and fp16/bf16).\n\n## Source\n\n- Repository: https://github.com/Blaizzy/mlx-audio\n- HuggingFace: https://huggingface.co/mlx-community\n"},"toolUseId":"toolu_0189M4gCkmdB8zkFsFYPXooY","output":"The file /home/pastilhas/projects/officer.dev/data/skills/mlxaudio/SKILL.md has been updated. Here's the result of running `cat -n` on a snippet of the edited file:\n     1→---\n     2→name: mlx.audio\n     3→description: Generate speech from text and transcribe audio using mlx-audio. Use when the user wants text-to-speech synthesis, speech-to-text transcription, voice cloning, audio separation, or speech-to-speech processing on Apple Silicon.\n     4→---\n     5→\n     6→# MLX-Audio\n     7→\n     8→A speech processing library built on Apple's MLX framework, providing TTS, STT, speech-to-speech (STS), and audio separation optimized for Apple Silicon.\n     9→\n    10→- **Repository:** https://github.com/Blaizzy/mlx-audio\n    11→- **License:** MIT\n    12→\n    13→## CLI Tools\n    14→\n    15→### Text-to-Speech (TTS)\n    16→\n    17→```bash\n    18→mlx_audio.tts.generate --model <model> --text '<text>' [options]\n    19→```\n    20→\n    21→| Flag | Type | Default | Description |\n    22→|------|------|---------|-------------|\n    23→| `--model` | string | required | HuggingFace model ID |\n    24→| `--text` | string | required | Text to synthesize |\n    25→| `--voice` | string | — | Voice preset (model-specific) |\n    26→| `--speed` | float | 1.0 | Speech speed multiplier |\n    27→| `--lang_code` | string | `a` | Language code |\n    28→| `--play` | flag | — | Play audio immediately |\n    29→| `--output_path` | string | — | Directory to save audio |\n    30→| `--ref_audio` | string | — | Reference audio for voice cloning (CSM) |\n    31→\n    32→#### Language Codes\n    33→\n    34→| Code | Language |\n    35→|------|----------|\n    36→| `a` | American English |\n    37→| `b` | British English |\n    38→| `j` | Japanese |\n    39→| `z` | Mandarin Chinese |\n    40→| `e` | Spanish |\n    41→| `f` | French |\n    42→\n    43→#### Kokoro Voices\n    44→\n    45→| Voice | Description |\n    46→|-------|-------------|\n    47→| `af_heart`, `af_bella`, `af_nova`, `af_sky` | American female |\n    48→| `am_adam`, `am_echo` | American male |\n    49→| `bf_alice`, `bf_emma` | British female |\n    50→| `bm_daniel`, `bm_george` | British male |\n    51→| `jf_alpha`, `jm_kumo` | Japanese |\n    52→| `zf_xiaobei`, `zm_yunxi` | Chinese |\n    53→\n    54→#### Examples\n    55→\n    56→```bash\n    57→# Basic generation\n    58→mlx_audio.tts.generate --model mlx-community/Kokoro-82M-bf16 --text 'Hello, world!' --lang_code a\n    59→\n    60→# With voice and speed\n    61→mlx_audio.tts.generate --model mlx-community/Kokoro-82M-bf16 --text 'Hello!' --voice af_heart --speed 1.2 --lang_code a\n    62→\n    63→# Play immediately\n    64→mlx_audio.tts.generate --model mlx-community/Kokoro-82M-bf16 --text 'Hello!' --play --lang_code a\n    65→\n    66→# Voice cloning with CSM\n    67→mlx_audio.tts.generate --model mlx-community/csm-1b --text \"Hello from Sesame.\" --ref_audio ./reference_voice.wav --play\n    68→```\n    69→\n    70→### Speech-to-Text (STT)\n    71→\n    72→```bash\n    73→python -m mlx_audio.stt.generate --model <model> --audio <file> [options]\n    74→```\n    75→\n    76→| Flag | Type | Default | Description |\n    77→|------|------|---------|-------------|\n    78→| `--model` | string | required | HuggingFace model ID |\n    79→| `--audio` | string | required | Input audio file |\n    80→| `--language` | string | — | Language code |\n    81→| `--max-tokens` | int | 1024 | Maximum output tokens |\n    82→| `--temperature` | float | 0.0 | Sampling temperature |\n    83→| `--context` | string | — | Hotwords/metadata for context |\n    84→| `--output-path` | string | — | Output directory |\n    85→| `--format` | string | — | Output format (e.g. `json`) |\n    86→| `--stream` | flag | — | Enable streaming mode |\n    87→| `--verbose` | flag | — | Detailed logging |\n    88→\n    89→#### Examples\n    90→\n    91→```bash\n    92→# Basic transcription\n    93→python -m mlx_audio.stt.generate --model mlx-community/whisper-large-v3-turbo-asr-fp16 --audio speech.wav --verbose\n    94→\n    95→# With context for technical terms\n    96→python -m mlx_audio.stt.generate --model mlx-community/VibeVoice-ASR-bf16 --audio meeting.wav --context \"MLX, Apple Silicon, PyTorch\" --max-tokens 8192 --format json --verbose\n    97→\n    98→# Parakeet model\n    99→python -m mlx_audio.stt.generate --model mlx-community/parakeet-tdt-0.6b-v3 --audio speech.wav --format json --verbose\n   100→```\n   101→\n   102→---\n   103→\n   104→## Python API\n   105→\n   106→### TTS\n   107→\n   108→```python\n   109→from mlx_audio.tts.utils import load_model\n   110→\n   111→model = load_model(\"mlx-community/Kokoro-82M-bf16\")\n   112→for result in model.generate(\"Hello from MLX-Audio!\", voice=\"af_heart\"):\n   113→    audio = result.audio  # mx.array waveform\n   114→```\n   115→\n   116→### STT\n   117→\n   118→```python\n   119→from mlx_audio.stt.generate import generate_transcription\n   120→\n   121→result = generate_transcription(\n   122→    model=\"mlx-community/whisper-large-v3-turbo-asr-fp16\",\n   123→    audio=\"audio.wav\",\n   124→)\n   125→print(result.text)\n   126→```\n   127→\n   128→### STT with Streaming\n   129→\n   130→```python\n   131→from mlx_audio.stt import load\n   132→\n   133→# VibeVoice-ASR streaming\n   134→model = load(\"mlx-community/VibeVoice-ASR-bf16\")\n   135→for text in model.stream_transcribe(audio=\"speech.wav\", max_tokens=4096):\n   136→    print(text, end=\"\", flush=True)\n   137→\n   138→# Parakeet streaming\n   139→model = load(\"mlx-community/parakeet-tdt-0.6b-v3\")\n   140→for chunk in model.generate(\"long_audio.wav\", stream=True):\n   141→    print(chunk.text, end=\"\", flush=True)\n   142→```\n   143→\n   144→### Forced Alignment (Qwen3)\n   145→\n   146→```python\n   147→from mlx_audio.stt import load\n   148→\n   149→aligner = load(\"mlx-community/Qwen3-ForcedAligner-0.6B-8bit\")\n   150→result = aligner.generate(\"audio.wav\", text=\"I have a dream\", language=\"English\")\n   151→for item in result:\n   152→    print(f\"[{item.start_time:.2f}s - {item.end_time:.2f}s] {item.text}\")\n   153→```\n   154→\n   155→---\n   156→\n   157→## REST API Server (OpenAI-compatible)\n   158→\n   159→### Starting the Server\n   160→\n   161→```bash\n   162→python -m mlx_audio.server [OPTIONS]\n   163→```\n   164→\n   165→| Flag | Type | Default | Description |\n   166→|------|------|---------|-------------|\n   167→| `--host` | string | `localhost` | Server host |\n   168→| `--port` | int | `8000` | Server port |\n   169→| `--allowed-origins` | string | `*` | CORS allowed origins |\n   170→| `--workers` | int/float | `2` | Number of workers |\n   171→| `--reload` | flag | — | Enable auto-reload |\n   172→| `--start-ui` | flag | — | Launch Studio UI alongside API |\n   173→| `--log-dir` | string | `logs` | Directory for server logs |\n   174→\n   175→### Endpoints\n   176→\n   177→#### GET /v1/models\n   178→\n   179→List available models.\n   180→\n   181→```bash\n   182→curl http://localhost:8000/v1/models\n   183→```\n   184→\n   185→#### POST /v1/models?model_name=\\<name\\>\n   186→\n   187→Add a model to the server.\n   188→\n   189→#### DELETE /v1/models?model_name=\\<name\\>\n   190→\n   191→Remove a model from the server.\n   192→\n   193→#### POST /v1/audio/speech\n   194→\n   195→Generate speech from text.\n   196→\n   197→```bash\n   198→curl -X POST http://localhost:8000/v1/audio/speech \\\n   199→  -H \"Content-Type: application/json\" \\\n   200→  -d '{\n   201→    \"model\": \"mlx-community/Kokoro-82M-bf16\",\n   202→    \"input\": \"Hello, world!\",\n   203→    \"voice\": \"af_heart\",\n   204→    \"speed\": 1.0,\n   205→    \"lang_code\": \"a\",\n   206→    \"response_format\": \"mp3\"\n   207→  }' --output speech.mp3\n   208→```\n   209→\n   210→| Parameter | Type | Default | Description |\n   211→|-----------|------|---------|-------------|\n   212→| `model` | string | required | Model ID |\n   213→| `input` | string | required | Text to synthesize |\n   214→| `voice` | string | — | Voice preset |\n   215→| `speed` | float | 1.0 | Speech speed |\n   216→| `lang_code` | string | `a` | Language code |\n   217→| `ref_audio` | string | — | Reference audio path (voice cloning) |\n   218→| `ref_text` | string | — | Reference transcript |\n   219→| `response_format` | string | `mp3` | Output format |\n   220→| `stream` | bool | false | Enable streaming |\n   221→| `streaming_interval` | float | 2.0 | Streaming chunk interval |\n   222→| `temperature` | float | 0.7 | Sampling temperature |\n   223→| `top_p` | float | 0.95 | Nucleus sampling |\n   224→| `top_k` | int | 40 | Top-k sampling |\n   225→| `repetition_penalty` | float | 1.0 | Repetition penalty |\n   226→| `max_tokens` | int | 1200 | Maximum tokens |\n   227→| `gender` | string | `male` | Gender hint |\n   228→| `pitch` | float | 1.0 | Pitch adjustment |\n   229→| `instruct` | string | — | Instruction text |\n   230→\n   231→#### POST /v1/audio/transcriptions\n   232→\n   233→Transcribe an audio file (multipart/form-data).\n   234→\n   235→```bash\n   236→curl -X POST http://localhost:8000/v1/audio/transcriptions \\\n   237→  -F file=@audio.wav \\\n   238→  -F model=mlx-community/whisper-large-v3-turbo-asr-fp16\n   239→```\n   240→\n   241→| Parameter | Type | Default | Description |\n   242→|-----------|------|---------|-------------|\n   243→| `file` | file | required | Audio file |\n   244→| `model` | string | required | Model ID |\n   245→| `language` | string | — | Language code |\n   246→| `max_tokens` | int | 1024 | Maximum tokens |\n   247→| `chunk_duration` | float | 30.0 | Chunk duration (seconds) |\n   248→| `stream` | bool | false | Enable streaming |\n   249→| `context` | string | — | Hotwords/context |\n   250→| `text` | string | — | Reference text |\n   251→| `verbose` | bool | false | Detailed output |\n   252→\n   253→Response (NDJSON stream):\n   254→\n   255→```json\n   256→{\"text\": \"chunk text\", \"accumulated\": \"full text so far\"}\n   257→```\n   258→\n   259→#### WebSocket /v1/audio/transcriptions/realtime\n   260→\n   261→Real-time transcription via WebSocket. Send initial config as JSON, then stream int16 PCM audio as binary frames.\n   262→\n   263→```json\n   264→{\n   265→  \"model\": \"mlx-community/whisper-large-v3-turbo-asr-fp16\",\n   266→  \"sample_rate\": 16000,\n   267→  \"streaming\": true\n   268→}\n   269→```\n   270→\n   271→#### POST /v1/audio/separations\n   272→\n   273→Separate audio sources (multipart/form-data).\n   274→\n   275→```bash\n   276→curl -X POST http://localhost:8000/v1/audio/separations \\\n   277→  -F file=@audio.wav \\\n   278→  -F model=mlx-community/sam-audio-large-fp16 \\\n   279→  -F description=\"speech\"\n   280→```\n   281→\n   282→| Parameter | Type | Default | Description |\n   283→|-----------|------|---------|-------------|\n   284→| `file` | file | required | Audio file |\n   285→| `model` | string | `mlx-community/sam-audio-large-fp16` | Model ID |\n   286→| `description` | string | `speech` | Target description |\n   287→| `method` | string | `midpoint` | ODE method (`midpoint` or `euler`) |\n   288→| `steps` | int | 16 | ODE steps (2/4/8/16/32) |\n   289→\n   290→Response:\n   291→\n   292→```json\n   293→{\n   294→  \"target\": \"<base64 WAV>\",\n   295→  \"residual\": \"<base64 WAV>\",\n   296→  \"sample_rate\": 44100\n   297→}\n   298→```\n   299→\n   300→---\n   301→\n   302→## Supported Models\n   303→\n   304→### TTS Models\n   305→\n   306→| Model | Languages | Notes |\n   307→|-------|-----------|-------|\n   308→| Kokoro | EN, JA, ZH, FR, ES, IT, PT, HI | Fast, high-quality multilingual |\n   309→| Qwen3-TTS | ZH, EN, JA, KO, + more | Voice design via instruction |\n   310→| CSM | EN | Voice cloning with reference audio |\n   311→| Dia | EN | Dialogue-focused |\n   312→| OuteTTS | EN | Efficient |\n   313→| Spark | EN, ZH | SparkTTS |\n   314→| Chatterbox | EN, ES, FR, DE, IT, PT, PL, TR, RU, NL, CS, AR, ZH, JA, HU, KO | Expressive multilingual |\n   315→| Soprano | EN | High-quality |\n   316→\n   317→### STT Models\n   318→\n   319→| Model | Languages | Notes |\n   320→|-------|-----------|-------|\n   321→| Whisper | 99+ languages | OpenAI's robust model |\n   322→| Qwen3-ASR | ZH, EN, JA, KO, + more | Alibaba multilingual |\n   323→| Qwen3-ForcedAligner | ZH, EN, JA, KO, + more | Word-level alignment |\n   324→| Parakeet | EN (v2), 25 EU languages (v3) | NVIDIA, high accuracy |\n   325→| Voxtral | Multiple | Mistral speech model |\n   326→| Voxtral Realtime | Multiple | 4B streaming STT |\n   327→| VibeVoice-ASR | Multiple | Microsoft 9B, supports diarization and context |\n   328→\n   329→### Other Models\n   330→\n   331→| Model | Type | Description |\n   332→|-------|------|-------------|\n   333→| Sortformer v1/v2.1 | VAD/Diarization | Speaker diarization (up to 4 speakers) |\n   334→| SAM-Audio | Separation | Text-guided source separation |\n   335→| Liquid2.5-Audio | STS | Speech/text-to-speech and STT |\n   336→| MossFormer2 SE | Enhancement | Speech enhancement / noise removal |\n   337→\n   338→Models are available from `mlx-community` on HuggingFace with various quantization levels (3-bit through 8-bit and fp16/bf16).\n   339→\n   340→## Source\n   341→\n   342→- Repository: https://github.com/Blaizzy/mlx-audio\n   343→- HuggingFace: https://huggingface.co/mlx-community\n   344→","isError":false}]