Speech recognition (ASR)

The ASR service transcribes audio to text using on-device models running fully on the Qualcomm NPU. It supports both file-based batch transcription and real-time streaming via WebSocket.

Prerequisites

Ensure the Audio Analytics container is deployed and running. See Running the Audio Analytics container for setup instructions.

Audio format requirements

Property	Requirement
Format	WAV only
Sample rate	16 kHz required. If your audio is a different sample rate, declare it via `parameters` in the create request.
Channels	Mono (auto-converted)
Max file size	25 MB

Available models

Query the models endpoint to get the list of available ASR models:

curl http://localhost:8085/audio-analytics/v1/api/transcriptions/models

Response:

[
  {
    "name": "whisper-small-quantized",
    "display_name": "Whisper Small (Quantized)",
    "description": "OpenAI whisper-small-quantized model for general transcription"
  }
]

Use the name value from this response in your transcription requests.

Endpoints

Method	Endpoint	Description
`GET`	`/transcriptions/models`	List available ASR models
`POST`	`/transcriptions/create`	Transcribe a file or start a streaming session
`POST`	`/transcriptions/close`	Close an active streaming session
`POST`	`/transcriptions/flush`	Force immediate processing of buffered audio

File transcription

Upload a WAV file and receive the complete transcription in a single response:

curl -X POST http://localhost:8085/audio-analytics/v1/api/transcriptions/create \
  -F "file=@recording.wav" \
  -F "model=whisper-small-quantized" \
  -F "language=en" \
  -F "stream=false"

Response:

{
  "text": "The transcribed text from your audio file.",
  "language": "en",
  "language_name": "English",
  "type": "transcript.text.done"
}

Set language to a language code supported by the model, or omit it for automatic detection.

File transcription with streaming results

Upload a complete WAV file and receive transcription results via WebSocket as the file is processed:

curl -X POST http://localhost:8085/audio-analytics/v1/api/transcriptions/create \
  -F "file=@recording.wav" \
  -F "model=whisper-small-quantized" \
  -F "language=en" \
  -F "stream=true"

The server acknowledges the upload and begins processing. Connect to the WebSocket to receive results. WebSocket results:

{ "state": "connection_established" }
{ "type": "transcript.text.done", "text": "The transcribed text.", "language": "en", "language_name": "English", "state": "transcription" }

Live audio streaming

Stream raw audio chunks in real time and receive transcription results as you speak.

Step 1 - Create a session

Declare the audio format via parameters if your audio is not 16 kHz mono PCM:

curl -X POST http://localhost:8085/audio-analytics/v1/api/transcriptions/create \
  -H "Content-Type: application/json" \
  -d '{
    "model": "whisper-small-quantized",
    "language": "en",
    "stream": true,
    "vad": 1200,
    "parameters": [
      {"key": "sampling_rate", "value": "44100"},
      {"key": "channels", "value": "1"},
      {"key": "format", "value": "pcm_s16le"}
    ]
  }'

Response:

{
  "session_id": "550e8400-e29b-41d4-a716-446655440000",
  "state": "asr_initialized",
  "text": "Successfully started Transcription Engine. Please connect to the WebSocket to send audio data & receive transcription output.",
  "language": "en",
  "type": "transcript.event"
}

The vad parameter sets the silence hangover in milliseconds before speech end is detected. Omit it to disable VAD and manage utterance end manually via /transcriptions/flush or /transcriptions/close.

Step 2 - Connect to WebSocket and send audio

Connect to ws://localhost:8085/stream. The server sends a connection_established message on connect before any transcription messages arrive:

{ "state": "connection_established" }

Send audio in 8192-byte chunks via WebSocket:

const ws = new WebSocket('ws://localhost:8085/stream');

ws.send(JSON.stringify({
  "message_type": "transcriptions_session_audio",
  "message_source": "audio_analytics_api",
  "session_id": "550e8400-e29b-41d4-a716-446655440000",
  "type": "input_audio",
  "data": "<BASE64_ENCODED_PCM>"
}));

Step 3 - Receive results

The server emits the following sequence per utterance:

{ "type": "transcript.event", "state": "speech_start", "session_id": "..." }
{ "type": "transcript.text.delta", "text": "Partial text...", "language": "en", "session_id": "..." }
{ "type": "transcript.event", "state": "speech_end", "session_id": "..." }
{ "type": "transcript.text.done", "text": "Complete utterance text.", "language": "en", "state": "transcription", "session_id": "..." }

After transcript.text.done the engine resets and listens for the next utterance. The session stays open until you call /transcriptions/close.

Step 4 - Close the session

curl -X POST http://localhost:8085/audio-analytics/v1/api/transcriptions/close \
  -H "Content-Type: application/json" \
  -d '{"session_id": "550e8400-e29b-41d4-a716-446655440000"}'

Response:

{
  "text": "",
  "language": null,
  "type": "asr_closed",
  "session_id": "550e8400-e29b-41d4-a716-446655440000",
  "state": "asr_closed"
}

Only one ASR session can be active at a time. Starting a second session while one is active returns HTTP 409 Conflict. Close the existing session first.

Voice Activity Detection (VAD)

VAD controls how the server detects the end of an utterance during a streaming session. When silence is detected for the configured duration, the server emits speech_end followed by transcript.text.done for the current utterance. The session remains open and resets for the next utterance — VAD does not close the session.

Parameter	Description
`vad`	Silence hangover in milliseconds before `speech_end` fires. For example, `1200` = 1.2 seconds of silence triggers end of utterance. Omit to disable VAD and manage utterance end manually via `/transcriptions/flush` or `/transcriptions/close`.

If no audio is received for the duration set by SESSION_TIMEOUT_S (default: 30 seconds) in the container config, the session closes automatically regardless of VAD state.

Flush buffered audio

Force immediate transcription of buffered audio without closing the session:

curl -X POST http://localhost:8085/audio-analytics/v1/api/transcriptions/flush \
  -H "Content-Type: application/json" \
  -d '{"session_id": "550e8400-e29b-41d4-a716-446655440000"}'

Python example

import requests

def transcribe_file(audio_file, language=None):
    url = "http://localhost:8085/audio-analytics/v1/api/transcriptions/create"
    data = {
        "model": "whisper-small-quantized",
        "stream": "false"
    }
    if language:
        data["language"] = language
    with open(audio_file, "rb") as f:
        response = requests.post(url, files={"file": f}, data=data, timeout=60)
    response.raise_for_status()
    return response.json()["text"]

print(transcribe_file("sample.wav"))

Troubleshooting

Poor transcription quality: Use 16 kHz mono WAV files and ensure audio is clear with minimal background noise.

IQ-9075 EVK

Prerequisites

Audio format requirements

Available models

Endpoints

File transcription

File transcription with streaming results

Live audio streaming

Step 1 - Create a session

Step 2 - Connect to WebSocket and send audio

Step 3 - Receive results

Step 4 - Close the session

Voice Activity Detection (VAD)

Flush buffered audio

Python example

Troubleshooting

Next steps

Text-to-speech

Language translation

​Prerequisites

​Audio format requirements

​Available models

​Endpoints

​File transcription

​File transcription with streaming results

​Live audio streaming

​Step 1 - Create a session

​Step 2 - Connect to WebSocket and send audio

​Step 3 - Receive results

​Step 4 - Close the session

​Voice Activity Detection (VAD)

​Flush buffered audio

​Python example

​Troubleshooting

​Next steps

Text-to-speech

Language translation

Prerequisites

Audio format requirements

Available models

Endpoints

File transcription

File transcription with streaming results

Live audio streaming

Step 1 - Create a session

Step 2 - Connect to WebSocket and send audio

Step 3 - Receive results

Step 4 - Close the session

Voice Activity Detection (VAD)

Flush buffered audio

Python example

Troubleshooting

Next steps