voice

name	voice
description	OminiX ASR (speech-to-text), preset-voice TTS with emotion/speed control, and model management via Qwen3 models on Apple Silicon. For voice cloning and custom voice profiles, use mofa-fm. Triggers: voice, transcribe audio, text to speech, speak this, read aloud, model management, download model, 语音识别, 语音合成, 模型管理.

OminiX ASR / TTS / Model Management

On-device speech-to-text, preset-voice text-to-speech with emotion control, and model management using OminiX Qwen3 ASR/TTS models on Apple Silicon.

Voice cloning and custom voice profiles are handled by mofa-fm (fm_tts, fm_voice_save, fm_voice_list, fm_voice_delete). This skill only supports preset voices.

Configuration

The skill auto-discovers the ominix-api server URL via (in priority order):

1. OMINIX_API_URL environment variable
2. Discovery file ~/.ominix/api_url (written by ominix-api on startup)
3. Default: http://localhost:9090

Checking Available Models

Use list_models to see what's installed. The response includes an endpoints array for each model, telling you which URL to use:

{"data": [
  {"id": "qwen3-asr", "type": "asr", "endpoints": ["/v1/audio/asr/qwen3"]},
  {"id": "Qwen3-TTS-CustomVoice-8bit", "type": "qwen3_tts", "endpoints": ["/v1/audio/tts/qwen3"]}
]}

If a model you need is missing, use download_model then load_model to install it.

API Endpoints

Function	Endpoint	Model
Preset TTS	POST /v1/audio/tts/qwen3	Qwen3-TTS CustomVoice
Qwen3-ASR	POST /v1/audio/asr/qwen3	Qwen3-ASR encoder-decoder
Paraformer	POST /v1/audio/asr/paraformer	Paraformer CTC-based

TTS and ASR run on separate threads — they do not block each other.

Tools

voice_transcribe

Transcribe an audio file to text via Qwen3-ASR. Supports WAV, OGG, MP3, FLAC, M4A.

{"audio_path": "/tmp/voice.ogg", "language": "Chinese"}

Parameters:

• audio_path (required): Absolute path to the audio file
• language (optional, default "Chinese"): "Chinese", "English", "Japanese", "Korean", "Cantonese", etc.

voice_synthesize

Generate speech audio from text using a preset voice. Uses Qwen3-TTS when ominix-api is running (high quality, emotion/style control). Falls back to macOS built-in say command when unavailable.

macOS Say auto-detects language from text and picks the appropriate built-in voice. Emotion prompts (prompt) are not supported in fallback mode.

{"text": "Hello world", "language": "english", "speaker": "ryan"}

With emotion:

{"text": "我太开心了！", "speaker": "vivian", "prompt": "用兴奋激动的语气说话，充满热情和活力"}

Parameters:

• text (required): Text to synthesize
• output_path (optional): Where to save audio. Default: auto-generated in OCTOS_WORK_DIR
• language (optional, default "chinese"): "chinese", "english", "japanese", "korean"
• speaker (optional, default "vivian"): Preset name only — vivian, serena, ryan, aiden, eric, dylan, uncle_fu, ono_anna, sohee
• prompt (optional): Style/emotion instruction (see tables below)
• speed (optional, default 1.0): Speed factor 0.5-2.0

Preset speakers: vivian, serena, ryan, aiden, eric, dylan (English/Chinese), uncle_fu (Chinese), ono_anna (Japanese), sohee (Korean)

Verified Chinese emotion prompts (best with vivian, serena, dylan, uncle_fu):

Style	Prompt
Excited	用兴奋激动的语气说话，充满热情和活力
Sad	用悲伤失望的语气说话，声音低沉，语速缓慢
Cheerful	用开朗愉快的语气说话，声音明亮上扬，节奏轻快
Shout	用大声喊叫的方式说话，声音高亢有力，语速快
Sarcastic	用讽刺嘲讽的语气说话，语调阴阳怪气，拖长尾音
Soft	用温柔轻柔的语气说话
Panic	用惊慌恐惧的语气说话，声音颤抖，语速急促

English emotion prompts (best with ryan, aiden):

Style	Prompt
Excited	Speak with excitement and enthusiasm, full of energy
Sad	Speak in a sad, disappointed tone, voice low and slow
Cheerful	Speak cheerfully with a bright, upbeat voice
Shout	Shout loudly with a powerful, high-pitched voice
Sarcastic	Speak sarcastically with a mocking, drawn-out tone
Soft	Speak gently and softly
Panic	Speak in a panicked, trembling voice, fast and breathless

Custom free-form prompts are also supported — include emotion + timbre + pace descriptors for strongest control.

list_models

List all loaded models and available models in the catalog.

download_model

Download a model from the catalog.

Parameters:

• model_id (required): Model ID from the catalog

load_model

Load a downloaded model into memory for inference.

Parameters:

• model (required): Model name or path
• model_type (optional, default "llm"): "llm", "asr", "tts"

unload_model

Unload a model from memory.

Parameters:

• model_type (required): Type of model to unload — "llm", "asr", "tts"