Navana Streaming ASR Instructions

How to Stream

Connection Instructions

• Endpoint: Websocket streaming speech API endpoint

  • wss://bodhi.navana.ai

• Sample Script:

  • streaming.py (for static audio files)

Access Token

Store the authentication headers in env to access the streaming speech API endpoints:

$ export API_KEY=YOUR_API_KEY
$ export CUSTOMER_ID=YOUR_CUSTOMER_ID

Brief Description of Response Format

The received response format will be a JSON object.

{
  "call_id": "CALL_ID",
  "segment_id": "SEGMENT_ID",
  "eos": false,
  "type": "partial",
  "text": "TRANSCRIPT"
}

Keys Description

• Call_id:

  • Unique identifier associated with every streaming connection

• Segment_id:

  • Unique identifier associated with every speech segment during the entire active socket connection

• Text:

  • If type = "partial"
    • Partial transcript corresponding to every streaming audio chunk
    • Partial transcripts for every audio chunk (will be for a 100ms audio chunk if streaming audio packet size is 100ms)
  • If type = "complete"
    • Complete/final transcript generated for each speech segment
    • Generated once per segment_id i.e., when the speech segment end is reached

• eos:

  • If 'eos' is true, marks the end of the streaming connection

Install packages

$ pip install -r requirements.txt

Usage

$ python streaming.py -f loan.wav

Options: -f: File name of the audio file to be streamed.

Audio Stream Requirements

To ensure optimal compatibility and performance with our audio processing system, please adhere to the following audio stream requirements:

• Encoding/Bit Depth: 16Bit PCM with a 2 Byte depth, providing high-quality audio representation.

• Minimum Sample Rate: The audio must have a sample rate of at least 8000Hz.

• Fixed Streaming Rate: Audio packets should be streamed at (chunk_duration_ms) a fixed size (50 - 500 ms), ensuring consistent data flow. We recommend using 100 ms as shown in the example script.

• Channels: Audio must be single-channel (Mono) to ensure compatibility with our processing pipeline.

• Speakers: Initially, support is provided for a single speaker per channel. However, support for multiple speakers on a single channel is under development and will be announced soon.

Available ASR Models for Testing

• Bengali: bn-general-jan24-v1-8khz
• Hindi: hi-general-feb24-v1-8khz
• Telugu: te-general-jan24-v1-8khz

For testing the code, modify the .py file with the model name you want to use.