Establishes a WebSocket connection for real-time audio transcription. This endpoint uses WebSocket protocol (wss://api.together.ai/v1/realtime) for bidirectional streaming communication.
Connection Setup:
Protocol: WebSocket (wss://)
Authentication: Pass API key as Bearer token in Authorization header
Parameters: Sent as query parameters (model, input_audio_format)
Client Events:
input_audio_buffer.append: Send audio chunks as base64-encoded data
input_audio_buffer.commit: Signal end of audio stream. When VAD is enabled, the server automatically detects speech boundaries and emits completed events. When VAD is disabled, you must send commit to trigger transcription of the buffered audio.
{ "type": "input_audio_buffer.commit"}
transcription_session.updated: Update session configuration, including Voice Activity Detection (VAD) parameters. Send this after receiving session.created. Can also be sent at any time during the session to change VAD settings.
VAD controls how the server automatically detects speech segments in the audio stream. When enabled (the default), the server uses Silero VAD to identify speech regions and emits transcription events as each segment completes. When disabled, you must manually call input_audio_buffer.commit to trigger transcription.
VAD can be configured in two ways:
Query parameters at connection time: turn_detection=server_vad&threshold=0.3&min_silence_duration_ms=500
Session message after connection: Send transcription_session.updated with a turn_detection object (see above)
To disable VAD at connection time, use turn_detection=none as a query parameter.
VAD Parameters:
All parameters are Omitted fields use their defaults.
Parameter
Type
Default
Description
type
string
server_vad
VAD mode. Use server_vad to enable, or set turn_detection to null to disable.
threshold
float
0.3
Speech probability threshold (0.0–1.0). Audio frames with probability above this value are classified as speech. Lower values detect more speech but may increase false positives. For low-SNR audio (e.g., 8kHz phone calls), values of 0.01–0.2 may work better.
min_silence_duration_ms
int
500
Minimum silence duration in milliseconds before ending a speech segment. Higher values merge nearby speech bursts into single segments. For phone calls with mid-sentence pauses, 2000–5000ms prevents over-segmentation.
min_speech_duration_ms
int
250
Minimum speech segment duration in milliseconds. Segments shorter than this are discarded. Filters out brief noise bursts or clicks.
max_speech_duration_s
float
5.0
Maximum speech segment duration in seconds. Segments longer than this are force-split at the longest internal silence gap. Useful for continuous speech without natural pauses.
speech_pad_ms
int
250
Padding in milliseconds added to the start and end of each detected segment. Prevents clipping speech edges. When padding would cause adjacent segments to overlap, the gap is split at the midpoint instead.
