Skip to content

Latest commit

 

History

History
92 lines (65 loc) · 3.43 KB

riva-asr-tts.md

File metadata and controls

92 lines (65 loc) · 3.43 KB
Raw

Optional: Enable NVIDIA Riva ASR and TTS

RAG Playground Integration

If you perform the following steps, you can use your voice to submit queries to the RAG Playground and the playground can read aloud the LLM responses.

Local Riva Server

To launch a Riva server locally, refer to the Riva Quick Start Guide.

  • In the provided config.sh script, make or confirm the following settings:

    • Set service_enabled_asr=true and service_enabled_tts=true.
    • Set the ASR and TTS languages by adding the language codes to asr_language_code and tts_language_code.
    • Set the models_tts variable to include a Rad-TTS model such as rmir_tts_radtts_hifigan_${modified_lang_code}_ipa.

Riva API Endpoint on NVIDIA API Catalog

You can access several GPU-accelerated speech models from NVIDIA API Catalog instead of hosting them locally. To access these models you need an API key. You can follow the steps to get an API key for the speech models.

Start the RAG Example

After you have access to a Riva server, perform the following steps to start any RAG example with speech support.

  1. Export the PLAYGROUND_MODE environment variable in your terminal:

    export PLAYGROUND_MODE=speech

  2. Edit the docker-compose.yaml file for the example and add the following environment variables to the rag-playground service:

    • Use speech models from NVIDIA API Catalog:

      rag-playgound:
  ...
  environment:
    RIVA_API_URI: grpc.nvcf.nvidia.com:443
    NVIDIA_API_KEY: ${NVIDIA_API_KEY}
    RIVA_ASR_FUNCTION_ID: 1598d209-5e27-4d3c-8079-4751568b1081 # nvidia/parakeet-ctc-riva-1-1b
    RIVA_TTS_FUNCTION_ID: 5e607c81-7aa6-44ce-a11d-9e08f0a3fe49 # nvidia/radtts-hifigan-riva
    TTS_SAMPLE_RATE: 48000

      To obtain your API key and function ID, navigate to the speech model and select Try API. From there, click Get API Key to generate your API key and refer Run python client to get function-id.

    • Use locally hosted RIVA models:

      rag-playgound:
  ...
  environment:
    RIVA_API_URI: <riva-ip-address>:50051
    TTS_SAMPLE_RATE: 48000

    • RIVA_API_URI is the Riva IP address or hostname and port number.

    • NVIDIA_API_KEY is the API Catalog key for a hosted Riva API endpoint.

    • RIVA_ASR_FUNCTION_ID is the Riva Function ID for a Riva API endpoint hosting ASR.

    • RIVA_TTS_FUNCTION_ID is the Riva Function ID for a Riva API endpoint hosting TTS.

    • TTS_SAMPLE_RATE is the text-to-speech sample rate in hertz.

  3. Start the containers:

    docker compose up -d --build

Next Steps

Refer to Using the Sample Web Application for information about enabling the Riva ASR and TTS features.