Fully offline Qwen3 TTS for your agent

High-quality text-to-speech using Qwen3-TTS. 10 built-in speakers with emotional instruct control, voice cloning (3s of audio), natural-language voice design, 10+ languages, persistent named voices, and delivering audio via Telegram/WhatsApp as native voice messages. Auto-detects GPU hardware (CUDA, ROCm, Intel XPU, CPU).

Qwen3-TTS Skill

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:

bash

bash <(curl -fsSL https://raw.githubusercontent.com/daMustermann/claw-qwen3-tts/main/install.sh)

Or if already cloned but not set up (no .venv/ directory):

bash

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:

bash

# Start (idempotent — won't restart if already running)
bash ~/clawd/skills/qwen3-tts/scripts/start_server.sh

# Check health
bash ~/clawd/skills/qwen3-tts/scripts/health_check.sh

# Stop (when done)
bash ~/clawd/skills/qwen3-tts/scripts/stop_server.sh

The server runs at http://localhost:8880.

---

Available Models

Model ID	Use Case	Notes
custom-voice-1.7b	High-quality TTS with built-in speakers — default	Best quality, ~5 GB VRAM
custom-voice-0.6b	Fast TTS with built-in speakers	Lightweight, ~2 GB VRAM
voice-design	Design new voices from natural language descriptions	Uses VoiceDesign model
base-1.7b	Basic TTS (auto-corrected to custom-voice-1.7b)	Use custom-voice-* instead
base-0.6b	Basic TTS (auto-corrected to custom-voice-0.6b)	Use custom-voice-* instead

Important: On the /v1/audio/speech endpoint, base-* and voice-design models are automatically corrected to the corresponding custom-voice-* model. Always prefer custom-voice-1.7b or custom-voice-0.6b for speech generation.

Built-in Speakers

The custom-voice-* models include 10 built-in voices:

Chelsie · Ethan · Aidan · Serena · Ryan · Vivian · Claire · Lucas · Eleanor · Benjamin

You can discover speakers dynamically: curl http://localhost:8880/v1/speakers

---

Capabilities

1. Generate Speech from Text

When to use: User asks to speak text, read something aloud, generate audio, do a voiceover, narrate, or say something.

bash

curl -X POST http://localhost:8880/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{
    "model": "custom-voice-1.7b",
    "input": "TEXT_HERE",
    "voice": "default",
    "speaker": "Chelsie",
    "language": "en",
    "instruct": "",
    "response_format": "wav"
  }' \
  --output ~/clawd/skills/qwen3-tts/output/speech.wav

Parameters:

Parameter	Required	Default	Description
model	no	custom-voice-1.7b	TTS model to use
input	yes	—	The text to synthesize
voice	no	default	"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.

bash

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.

bash

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:

1. 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?"

2. If the user says yes, capture the X-Voice-Id from the response headers and save it:

   bash

   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"
     }'
   

3. 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

4. When user asks to list voices:

   bash

   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).

5. When user asks to delete a voice: Confirm with the user first, then:

   bash

   curl -X DELETE http://localhost:8880/v1/voices/VOICE_NAME
   

6. When user asks to rename a voice:

   bash

   curl -X PATCH http://localhost:8880/v1/voices/OLD_NAME \
     -H "Content-Type: application/json" \
     -d '{"name": "NEW_NAME"}'
   

7. When user asks to update a voice's metadata (description, tags, language):

   bash

   curl -X PATCH http://localhost:8880/v1/voices/VOICE_NAME \
     -H "Content-Type: application/json" \
     -d '{"description": "Updated description", "tags": ["new", "tags"]}'
   

8. Voice names are case-insensitive but stored in the casing the user provided.

9. 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.

10. 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.

bash

curl -X POST http://localhost:8880/v1/audio/convert \
  -F "audio=@input.wav" \
  -F "target_format=mp3" \
  --output output.mp3

Supported formats: wav, mp3, ogg (Opus), flac

You can also use the shell script directly:

bash

bash ~/clawd/skills/qwen3-tts/scripts/convert_to_ogg_opus.sh input.wav output.ogg

6. Send via Telegram (PTT Voice Message)

When to use: User is interacting via Telegram, or explicitly asks to send audio to a Telegram chat.

bash

curl -X POST http://localhost:8880/v1/audio/send/telegram \
  -H "Content-Type: application/json" \
  -d '{
    "audio_file": "/path/to/audio.wav",
    "chat_id": "CHAT_ID",
    "bot_token": "BOT_TOKEN",
    "caption": "Optional caption"
  }'

• bot_token is optional if already configured in config.json
• Audio is auto-converted to OGG/Opus and sent via Telegram's sendVoice API
• Displays as a native PTT waveform voice message in the chat

7. Send via WhatsApp (PTT Voice Message)

When to use: User is interacting via WhatsApp, or explicitly asks to send audio there.

bash

curl -X POST http://localhost:8880/v1/audio/send/whatsapp \
  -H "Content-Type: application/json" \
  -d '{
    "audio_file": "/path/to/audio.wav",
    "phone_number_id": "PHONE_ID",
    "recipient": "[phone]",
    "access_token": "ACCESS_TOKEN"
  }'

• 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:

bash

# 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:

1. Tell the user the audio has been generated
2. Provide the output file path
3. If it was voice-design or voice-clone, always ask to save the voice (Rule §4.1)
4. 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
• Default audio format: wav, mp3, ogg, flac
• Device override: auto, cuda:0, xpu:0, cpu

If config.json doesn't exist, copy the template:

bash

cp ~/clawd/skills/qwen3-tts/config.json.template ~/clawd/skills/qwen3-tts/config.json