sarvam-transcribe-streaming-sdk 0.1.0

Python SDK for Sarvam Streaming Speech-to-Text

Sarvam Streaming Speech-to-Text — Python SDK

Real-time speech-to-text for Indian languages. Stream audio in, get transcripts back — with sentence-level segmented transcripts for high-quality, complete outputs.

Installation

pip install sarvam-transcribe-streaming-sdk

Requires Python 3.9+.

Quick Start

import asyncio
from sarvam_transcribe import SarvamStreamingClient
from sarvam_transcribe.handlers import TranscriptResultStreamHandler
from sarvam_transcribe.model import TranscriptEvent

SARVAM_STT_URL = "wss://stt.example.com"
SARVAM_API_KEY = "your-api-key"
AUDIO_FILE = "recording.pcm"    # 16kHz, mono, 16-bit PCM
CHUNK_SIZE = 16000               # 500ms of audio


class MyHandler(TranscriptResultStreamHandler):
    async def handle_transcript_event(self, event: TranscriptEvent):
        for result in event.transcript.results:
            for alt in result.alternatives:
                prefix = "..." if result.is_partial else ">>>"
                print(f"{prefix} {alt.transcript}")


async def main():
    client = SarvamStreamingClient(
        url=SARVAM_STT_URL,
        api_key=SARVAM_API_KEY,
    )
    stream = await client.start_stream_transcription(
        language_code="hi-IN",
    )

    async def send_audio():
        with open(AUDIO_FILE, "rb") as f:
            while chunk := f.read(CHUNK_SIZE):
                await stream.input_stream.send_audio_event(audio_chunk=chunk)
                await asyncio.sleep(CHUNK_SIZE / (16000 * 2))  # real-time pacing
        await stream.input_stream.end_stream()

    handler = MyHandler(stream.output_stream)
    await asyncio.gather(send_audio(), handler.handle_events())


asyncio.run(main())

Output:

... प्रधानमंत्री ने
... प्रधानमंत्री ने आज कहा कि
>>> प्रधानमंत्री ने आज कहा कि देश आगे बढ़ रहा है।
... हमें
... हमें मिलकर काम
>>> हमें मिलकर काम करना होगा।

Lines prefixed with ... are partial results (the transcript so far, which may still change). Lines prefixed with >>> are final results (complete, endpointed sentences ready for downstream use).

Supported Languages

Language	Code
Hindi	hi-IN
Bengali	bn-IN
Kannada	kn-IN
Malayalam	ml-IN
Marathi	mr-IN
Odia	od-IN
Punjabi	pa-IN
Tamil	ta-IN
Telugu	te-IN
English (Indian)	en-IN
Gujarati	gu-IN
Auto-detect	unknown

Audio Requirements

Property	Value
Format	Raw PCM (no WAV header)
Sample rate	16,000 Hz
Bit depth	16-bit signed, little-endian
Channels	Mono (1 channel)
Recommended chunk size	100–500 ms (3,200–16,000 bytes)
Max chunk size	32 KB (1 second)

Documentation

For full API reference, examples (microphone streaming, file transcription, collecting results), error handling, and migration guide from AWS Transcribe SDK, contact us.

License

Apache-2.0