Coqui TTS is an open-source, high-quality text-to-speech engine built on deep learning.
It provides natural-sounding voices and supports multiple languages, making it a great option for local deployments where flexibility and privacy are priorities.
¶ Coqui TTS Integration
AVR integrates with CoquiTTS to enable real-time speech synthesis inside your telephony infrastructure.
¶ Advantages of Using CoquiTTS
- Open Source → No vendor lock-in, free to customize
- Local Deployment → Runs fully on your own hardware, ideal for privacy-sensitive use cases
- Multi-language Support → Large selection of pre-trained voice models available
- Flexible Models → Choose from different architectures (VITS, FastPitch, Tacotron2, etc.)
- Custom Voices → Train your own voices with open datasets
¶ Configuration
¶ Environment Variables
| Variable | Description | Example Value |
|---|---|---|
PORT |
Port where the CoquiTTS service runs | 6032 |
Example .env file:
PORT=6032
¶ Docker Compose Example
Below is a sample configuration using AVR with CoquiTTS:
avr-tts-coquitts:
image: agentvoiceresponse/avr-tts-coquitts
platform: linux/x86_64
container_name: avr-tts-coquitts
restart: always
environment:
- PORT=6032
- COQUI_AI_TTS_URL=http://avr-coqui-ai-tts:5002/api/tts
ports:
- 6032:6032
networks:
- avr
avr-coqui-ai-tts:
image: ghcr.io/coqui-ai/tts-cpu
platform: linux/x86_64
container_name: avr-coqui-ai-tts
entrypoint: "python3 TTS/server/server.py --model_name tts_models/en/vctk/vits"
restart: always
environment:
- PORT=5002
ports:
- 5002:5002
networks:
- avr
Notes:
- avr-tts-coquitts acts as the AVR microservice wrapper.
- avr-coqui-ai-tts runs the official Coqui TTS server with a selected model.
- You can replace tts_models/en/vctk/vits with any supported model from Coqui TTS models.
¶ How It Works
- AVR sends text requests to avr-tts-coquitts.
- The request is forwarded to the running Coqui TTS server (avr-coqui-ai-tts).
- Coqui synthesizes speech using the selected model.
- The output is downsampled to 8kHz PCM for compatibility with Asterisk AudioSocket.
- AVR streams audio back to the caller in 320-byte chunks for real-time playback.
¶ Example Test with cURL
curl -X POST http://localhost:6032/text-to-speech-stream \
-H "Content-Type: application/json" \
-d '{"text": "Hello, this is CoquiTTS running with Agent Voice Response!"}' \
--output response.raw
The resulting response.raw file will contain PCM audio at 8kHz.
¶ Best Practices
- Test different Coqui models (tts_models/...) to find the best trade-off between latency and voice quality.
- Run with GPU acceleration for production or real-time scenarios.
- Keep your models directory mounted as a Docker volume to avoid re-downloading on container restarts.
¶ Community & Support
- GitHub: avr-tts-coquitts