Streaming STT

This guide explains how to implement streaming speech-to-text. Two protocols are supported: 1) gRPC and 2) WebSocket. See Streaming STT - gRPC and Streaming STT - WebSocket for integration details. For file-based transcription, see Batch STT.

caution

Streaming STT is subject to concurrent channel limits. See the rate limit policy.

The server may close a streaming connection if no audio chunk is sent within 60 seconds after the connection is established.

Audio Encoding

Streaming STT supports LINEAR16, FLAC, MULAW, ALAW, AMR, AMR_WB, OGG_OPUS, and OPUS.

caution

• Send LINEAR16, MULAW, ALAW, AMR, and AMR_WB as raw audio data.
• Send OGG_OPUS as OPUS-encoded audio frames in an OGG container.
• OPUS is supported only over gRPC. Contact us if you need to use it.

Common DecoderConfig/Parameter Information

The following DecoderConfig fields are shared by gRPC and WebSocket integrations.

Name	Type (gRPC / WebSocket)	Description	Required	Default
sample_rate	int	Range: 8000–48000 Hz	Yes	-
encoding	AudioEncoding / string	See Audio Encoding	Yes	-
model_name	string	Recognition model: sommers_ko (Korean), sommers_ja (Japanese), or whisper (Korean fine-tuned, multilingual)	No	sommers_ko
domain	string	See Domain	No	CALL
use_itn	bool	See Inverse Text Normalization	No	true
use_disfluency_filter	bool	See Disfluency Filter	No	false
use_profanity_filter	bool	See Profanity Filter	No	false
use_punctuation	bool	Whether to add punctuation	No	false
keywords	string[] / string	See Keywords Boosting	No	-
language	string	Used with the whisper model. See Recognition Language	No	ko

Keywords Boosting

Keywords boosting increases or decreases the recognition probability of specific words in Streaming STT.

caution

This feature is currently available only with the sommers_ko model or the whisper model when language is set to ko.

Format and Usage

Specify each keyword and an optional score:

• gRPC accepts a string[].
• WebSocket accepts a single comma-separated string.

Each keyword must use one of the following formats:

• "word": uses the default score of 2.0
• "word:score": separates the word and score with a colon

Keywords Boosting Rules

• Word: Use only Hangul syllables and spaces.
• Score:
  • Range: -5.0 to 5.0
  • A positive value increases recognition probability.
  • A negative value decreases recognition probability.
  • A value of 0 applies no score. Remove keywords that do not need boosting because their presence can still affect other keywords.

Example

// gRPC
["부스팅", "리턴제로:3.5", "에스티티:-1"]

// WebSocket
"부스팅,리턴제로:3.5,에스티티:-1"

caution

• Scores must be between -5.0 and 5.0.
• Write words according to their Korean pronunciation. For example, use 에스티티 instead of STT.
• Words must contain only Hangul syllables and spaces.
• Each word is limited to 20 characters, with up to 100 words.
• Registering a longer word that starts with a shorter keyword can reduce the boosting effect of the shorter word. For example, this can occur when registering both 커피 and 커피숍.

AI Voice Agent

When an AI voice agent or callbot detects the end of a user's utterance with its own VAD or turn detector, it can use the FINALIZE command to finalize the current utterance. This lets the client receive the final transcript and start generating the next response without waiting for server-side end-point detection.

FINALIZE does not close the streaming connection. After receiving the final transcript, continue sending audio for the next user turn over the same connection. To end the session, half-close the gRPC request stream or send the EOS text message over WebSocket. An empty utterance or repeated FINALIZE commands produce no result. See the protocol-specific request examples for gRPC and WebSocket.

caution

For container-based encodings such as OGG_OPUS, start the next utterance as an independent stream with a new container header after FINALIZE.

Domain

The domain optimizes the STT model for the user's audio input environment.

Supported Domains

• CALL (default): Optimized for phone calls and environments where the microphone is close to the speaker.
  • Clearly recognizes the nearby speaker.
  • Minimizes background noise and other voices.
  • Suitable for phone calls, personal headsets, and close-talk microphones.
• MEETING: Optimized for far-field audio in environments such as meeting rooms.
  • Handles speech from multiple speakers.
  • Suitable for meeting rooms and public spaces with distant microphones.

Choose the domain that matches the actual audio input environment.