For AI agents: a documentation index is available at the root level at /llms.txt. Append /llms.txt to any URL for a page-level index, or .md for the markdown version of any page.
WebSocket channel for real-time speech to text streaming powered by the
`saaras:v3-realtime` model.
**Authentication:** Pass your API subscription key via the
`API-SUBSCRIPTION-KEY` header, or (in browsers) via the WebSocket
subprotocol `api-subscription-key.<key>` — the server echoes the
subprotocol back.
**Note:** This API Reference page is provided for informational purposes only.
The Try It playground may not provide the best experience for streaming audio.
For optimal streaming performance, please use the SDK or implement your own WebSocket client.
WebSocket channel for real-time speech to text streaming powered by the
saaras:v3-realtime model.
Authentication: Pass your API subscription key via the
API-SUBSCRIPTION-KEY header, or (in browsers) via the WebSocket
subprotocol api-subscription-key.<key> — the server echoes the
subprotocol back.
Note: This API Reference page is provided for informational purposes only.
The Try It playground may not provide the best experience for streaming audio.
For optimal streaming performance, please use the SDK or implement your own WebSocket client.
Handshake
WSS
wss://api.sarvam.ai/speech-to-text-realtime/ws
Headers
Api-Subscription-KeystringRequired
API subscription key for authentication
Query parameters
language_codeenumRequired
BCP-47 language code of the input audio. **Required.**
**Supported languages (saaras:v3-realtime):**
- `auto`: Adaptive automatic language detection
- `en-IN`: English
- `hi-IN`: Hindi
- `bn-IN`: Bengali
- `kn-IN`: Kannada
- `ml-IN`: Malayalam
- `mr-IN`: Marathi
- `or-IN`: Odia
- `pa-IN`: Punjabi
- `ta-IN`: Tamil
- `te-IN`: Telugu
- `gu-IN`: Gujarati
- `as-IN`: Assamese
- `ur-IN`: Urdu
- `ne-IN`: Nepali
- `kok-IN`: Konkani
- `ks-IN`: Kashmiri
- `sd-IN`: Sindhi
- `sa-IN`: Sanskrit
- `sat-IN`: Santali
- `mni-IN`: Manipuri
- `brx-IN`: Bodo
- `mai-IN`: Maithili
- `doi-IN`: Dogri
modelenumOptionalDefaults to saaras:v3-realtime
Speech-to-text model. Only saaras:v3-realtime is accepted on this endpoint.
Allowed values:
stream_typeenumOptionalDefaults to balanced
Controls audio chunking and latency behaviour.
- **fast**: Partials at low latency.
- **balanced** (default): Partials with better accuracy.
- **simulated**: No partials.
Allowed values:
modeenumOptionalDefaults to transcribe
Task applied on the final transcript, not on partial transcripts.
Streaming partials are always straight transcription regardless of
this value.
- **transcribe** (default): Standard transcription in the source language script.
- **translate**: Translation to English.
- **verbatim**: Exact transcription in the original script (includes disfluencies).
- **translit**: Transliteration to Latin script.
- **codemix**: Code-mixed output (native + English).
Allowed values:
promptstringOptional
Optional context/terminology hint to bias decoding. Applied on the
final transcript.
Audio sample rate in Hz. Must match the actual audio.
**Accepted values:**
- `8000`
- `16000` (default)
Any other value closes the connection with code `4000`.
Allowed values:
return_timestampsenumOptionalDefaults to false
When true, adds segment-level start_s / end_s audio-time offsets
to transcript.final.
Allowed values:
thresholdstringOptionalDefaults to 0.3
Speech-detection sensitivity (0.0–1.0; higher = needs louder/clearer
speech). Only applies when endpointing=vad.
prefix_padding_msstringOptionalDefaults to 300
Audio (ms, ≥0) included before the detected speech onset. Only applies
when endpointing=vad.
silence_duration_msstringOptionalDefaults to 500
Silence (ms, ≥0) after speech that marks end-of-turn. Only applies when
endpointing=vad.
min_speech_duration_msstringOptionalDefaults to 250
Minimum speech duration (ms, ≥0) for a valid utterance. Only applies
when endpointing=vad.
Send
Realtime Audio InputobjectRequired
Send an audio chunk for real-time transcription (saaras:v3-realtime)
OR
Realtime Speech StartobjectRequired
Signal the start of an utterance (endpointing=manual only)
OR
Realtime Speech EndobjectRequired
Signal the end of an utterance (endpointing=manual only)
OR
Realtime FlushobjectRequired
OR
Realtime Config UpdateobjectRequired
Update session config mid-stream
OR
Realtime End SessionobjectRequired
Gracefully end the realtime session
OR
Realtime PingobjectRequired
Send a keepalive ping to the realtime WebSocket
Receive
Realtime Session BeginobjectRequired
Receive the session.begin event with the resolved session config
OR
Realtime VAD EventsobjectRequired
Receive VAD speech start/end events (endpointing=vad only)
OR
Realtime VAD EventsobjectRequired
Receive VAD speech start/end events (endpointing=vad only)
OR
Realtime Partial TranscriptobjectRequired
Receive streaming partial transcripts during an utterance
OR
Realtime Final TranscriptobjectRequired
Receive the final transcript for a completed utterance
OR
Realtime Config UpdatedobjectRequired
Receive the config.updated acknowledgement after a config.update
OR
Realtime PongobjectRequired
Receive the pong response to a client ping
OR
Realtime Session EndobjectRequired
Receive the session.end summary when the session closes
OR
Realtime ErrorobjectRequired
Receive non-fatal or fatal error notifications
BCP-47 language code of the input audio. Required.
Supported languages (saaras:v3-realtime):
auto: Adaptive automatic language detection
en-IN: English
hi-IN: Hindi
bn-IN: Bengali
kn-IN: Kannada
ml-IN: Malayalam
mr-IN: Marathi
or-IN: Odia
pa-IN: Punjabi
ta-IN: Tamil
te-IN: Telugu
gu-IN: Gujarati
as-IN: Assamese
ur-IN: Urdu
ne-IN: Nepali
kok-IN: Konkani
ks-IN: Kashmiri
sd-IN: Sindhi
sa-IN: Sanskrit
sat-IN: Santali
mni-IN: Manipuri
brx-IN: Bodo
mai-IN: Maithili
doi-IN: Dogri
Controls audio chunking and latency behaviour.
fast: Partials at low latency.
balanced (default): Partials with better accuracy.
simulated: No partials.
Task applied on the final transcript, not on partial transcripts.
Streaming partials are always straight transcription regardless of
this value.
transcribe (default): Standard transcription in the source language script.
translate: Translation to English.
verbatim: Exact transcription in the original script (includes disfluencies).
translit: Transliteration to Latin script.
codemix: Code-mixed output (native + English).
Turn detection strategy.
vad (default): Voice Activity Detection automatically detects turn boundaries.
manual: The client delimits turns with speech_start / speech_end events.
Audio sample rate in Hz. Must match the actual audio.
Accepted values:
8000
16000 (default)
Any other value closes the connection with code 4000.
