Skip to content

enable-zero-shot-tts.md

Latest commit

 

History

History
55 lines (42 loc) · 2.02 KB

enable-zero-shot-tts.md

File metadata and controls

55 lines (42 loc) · 2.02 KB

Enable Zero-Shot TTS

Zero-shot TTS allows you to clone any voice using a short audio sample (5+ seconds). This feature uses the Magpie Zero-shot model. Apply for access at the NVIDIA RIVA TTS Zero-shot models page.

Steps

  1. Prepare your audio sample to meet the following requirements.

    • Audio format is WAV (16-bit PCM recommended).
    • Audio duration is at least 5 seconds of clean speech.
    • Audio quality is clear, no background noise.
    • Language should match your target language.
    • Speaker is single speaker only.

  2. Create the audio_prompts directory and add your voice sample.

    mkdir -p audio_prompts
cp <your_audio_file>.wav audio_prompts/<your_audio_file>.wav

  3. Run the Magpie zero-shot NIM microservice following the instructions in the NVIDIA NIM RIVA TTS documentation.

  4. Set the environment variables in the .env file as follows.

    # Comment out standard TTS configuration
#TTS_DOCKER_IMAGE=nvcr.io/nim/nvidia/magpie-tts-multilingual:1.6.0
#TTS_VOICE_ID=Magpie-Multilingual.EN-US.Aria
#TTS_MODEL_NAME=magpie_tts_ensemble-Magpie-Multilingual
#TTS_NIM_TAGS=name=magpie-tts-multilingual,batch_size=32

# Enable Zero-shot TTS
TTS_DOCKER_IMAGE=<ZEROSHOT_NIM_MICROSERVICE_IMAGE> # Use your version
TTS_VOICE_ID=Magpie-ZeroShot.Female-1
TTS_MODEL_NAME=magpie_tts_ensemble-Magpie-ZeroShot
TTS_NIM_TAGS=name=magpie-tts-zeroshot,batch_size=32
ZERO_SHOT_AUDIO_PROMPT=audio_prompts/custom_voice.wav

  5. Update the python-app service in docker-compose.yml to mount audio prompts:

    python-app:
  # ... existing configuration ...
  volumes:
    - ./audio_dumps:/app/audio_dumps
    - ./config/:/app/config/
    - ./audio_prompts:/app/audio_prompts  # Add this line

  6. Deploy the services:

    docker compose up -d