Instalar com Codex ou Claude Copie este prompt, cole no Codex, Claude ou outro assistente e deixe que ele revise a página da skill e instale para você.
Um comando direto ignora o prompt de revisão. Verifique a origem antes de executá-lo.
You have access to a powerful text-to-speech system that can generate human-quality speech with 10 built-in speakers, design new voices from descriptions, clone existing voices from audio samples, and send audio via Telegram/WhatsApp as native voice messages.
First-Time Setup
If the skill is not yet installed (no ~/clawd/skills/qwen3-tts directory), run:
Or if already cloned but not set up (no .venv/ directory):
bash ~/clawd/skills/qwen3-tts/install.sh
This auto-detects the GPU (CUDA, ROCm, Intel XPU, or CPU-only), creates a Python venv, and installs all dependencies. It takes 5–15 minutes on first run.
Starting & Stopping the Server
Before any TTS operation, ensure the server is running:
"default" for built-in speakers, or a saved voice name (e.g. "Angie")
speaker
no
Chelsie
Built-in speaker name (only when voice is "default")
language
no
en
Language code: en, zh, ja, ko, de, fr, ru, pt, es, it
instruct
no
""
Emotional/style instruction (see below)
response_format
no
wav
Output format: wav, mp3, ogg, flac
speed
no
1.0
Speech speed multiplier
Language codes:en, zh, ja, ko, de, fr, ru, pt, es, it — or full names like English, Chinese, German, etc.
Instruct examples (controls tone, emotion, and style):
"Speak happily and with excitement"
"Whisper softly, as if telling a secret"
"Read this in a calm, professional news anchor tone"
"用愤怒的语气" (Speak angrily — works in target language too)
"" (empty string = neutral default)
When voice is a saved name: If you pass "voice": "Angie" and a voice named "Angie" exists, the server uses voice cloning with the saved reference audio instead of a built-in speaker. The speaker field is ignored in this case.
2. Design a New Voice
When to use: User wants to create a custom voice, describe how a character should sound, design a persona's voice.
curl -X POST http://localhost:8880/v1/audio/voice-design \
-H "Content-Type: application/json" \
-d '{
"model": "voice-design",
"input": "TEXT_TO_SPEAK",
"voice_description": "DESCRIBE THE VOICE IN NATURAL LANGUAGE",
"language": "en",
"response_format": "wav"
}' \
--output ~/clawd/skills/qwen3-tts/output/designed.wav
Parameters:
Parameter
Required
Default
Description
model
no
voice-design
Must be voice-design
input
yes
—
Text to synthesize with the designed voice
voice_description
yes
—
Natural language description of the desired voice
language
no
en
Target language
response_format
no
wav
Output format
Example descriptions:
"A warm, deep male voice with a slight British accent, calm and authoritative, like a BBC presenter in his 40s"
"A young, energetic female voice, bright and cheerful, with a slight rasp"
"An old wizard with a slow, mysterious, gravelly voice"
The response includes a X-Voice-Id header — capture it to save the voice (see §4).
3. Clone a Voice
When to use: User provides a reference audio clip and wants to generate new speech in that voice.
curl -X POST http://localhost:8880/v1/audio/voice-clone \
-F "reference_audio=@/path/to/reference.wav" \
-F "reference_text=Transcript of the reference audio" \
-F "input=New text to speak in the cloned voice" \
-F "language=en" \
-F "response_format=wav" \
--output ~/clawd/skills/qwen3-tts/output/cloned.wav
Parameters:
Parameter
Required
Default
Description
reference_audio
yes
—
Audio file to clone the voice from
input
yes
—
New text to synthesize in the cloned voice
reference_text
no
""
Transcription of the reference audio (improves quality)
language
no
en
Target language
response_format
no
wav
Output format
Guidelines:
Minimum 3 seconds of reference audio
Recommended 10–30 seconds for best quality
Providing an accurate reference_text transcription significantly improves results
Supports cross-language cloning (clone from English → speak in Japanese)
If reference_text is empty, uses x-vector-only mode (audio features only)
The response includes a X-Voice-Id header — capture it to save the voice (see §4).
4. ⭐ CRITICAL: Voice Save Prompting Rules
YOU MUST FOLLOW THESE RULES:
After EVERY voice-design or voice-clone request, ask the user:
"Would you like to save this voice for future use? What name should I give it?"
If the user says yes, capture the X-Voice-Id from the response headers and save it:
curl -X POST http://localhost:8880/v1/voices \
-H "Content-Type: application/json" \
-d '{
"name": "USER_CHOSEN_NAME",
"source_voice_id": "VOICE_ID_FROM_X_VOICE_ID_HEADER",
"description": "Description of the voice",
"tags": ["tag1", "tag2"],
"language": "en"
}'
When user requests TTS with a voice name (e.g. "say this with Angie"):
Use "voice": "Angie" in the /v1/audio/speech request
The server automatically loads the saved reference audio and uses voice cloning
If the name doesn't exist, tell the user and offer to design or clone one
When user asks to list voices:
curl http://localhost:8880/v1/voices
Present the results as a formatted list with name, description, source, language, tags, and usage count. Voices are sorted by usage count (most used first).
When user asks to delete a voice: Confirm with the user first, then:
Voice names are case-insensitive but stored in the casing the user provided.
No duplicate names allowed. If a name already exists, the save will fail (409). Ask the user for a different name or offer to delete the existing one first.
Voice profiles are stored locally in ~/clawd/skills/qwen3-tts/voices/ and persist across server restarts. Each voice consists of:
<name>.json — metadata
<name>.pt — embedding tensor
<name>_sample.wav — reference audio sample (used for re-cloning)
5. Convert Audio Formats
When to use: User needs audio in a specific format, or you need to prepare audio for messaging.
phone_number_id and access_token are optional if already configured in config.json
Audio is auto-converted to OGG/Opus and sent as a native WhatsApp voice message
8. Discovery Endpoints
Use these to dynamically discover available models and speakers:
# List all available TTS models
curl http://localhost:8880/v1/models
# List built-in speakers
curl http://localhost:8880/v1/speakers
# Server health check (device info, voice count, version)
curl http://localhost:8880/health
How to Respond
After generating speech:
Tell the user the audio has been generated
Provide the output file path
If it was voice-design or voice-clone, always ask to save the voice (Rule §4.1)
If the user is on Telegram/WhatsApp, offer to send it as a voice message
After saving a voice:
Confirm the name and tell the user they can use it anytime with that name
Example: "Voice saved as 'Captain Hook'! You can reference it anytime with voice: Captain Hook."
After sending via Telegram/WhatsApp:
Confirm successful delivery
When choosing a speaker: If the user doesn't specify, default to "Chelsie". If they describe the kind of voice they want (but not a full voice-design request), pick the most fitting built-in speaker.
When choosing a model: Default to custom-voice-1.7b. Only use custom-voice-0.6b if the user asks for speed, or if the system has limited VRAM/memory.
Configuration
The agent can update ~/clawd/skills/qwen3-tts/config.json to set:
Telegram: bot token and default chat ID
WhatsApp: phone number ID and access token
Default model:custom-voice-1.7b or custom-voice-0.6b
