Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/huggingface/speech-to-speech/llms.txt

Use this file to discover all available pages before exploring further.

VADHandlerArguments controls the Silero VAD v5 stage that sits at the front of every pipeline mode. The VAD listens to the audio stream continuously, segments it into speech chunks, and gates when STT and LLM processing begins. Tuning these flags has a direct impact on end-to-end latency, barge-in sensitivity, and false-positive suppression. All fields are passed without a prefix, for example --thresh 0.5 or --min_silence_ms 100.

​
Fields

​
thresh
float
default:"0.6"
Confidence threshold above which Silero VAD declares a frame as speech. Values range from 0.0 to 1.0. Lower values increase sensitivity (more detections, potentially more false positives); higher values require stronger evidence of speech before triggering.
speech-to-speech --thresh 0.5
​
sample_rate
integer
default:"16000"
Expected sample rate of the incoming audio in Hertz. The pipeline records and resamples all audio to 16 000 Hz before passing it to the VAD, so this should stay at its default unless you are providing audio at a different rate from an external source.
speech-to-speech --sample_rate 16000
​
min_silence_ms
integer
default:"64"
Minimum length of a silence interval (in milliseconds) that causes the VAD to end the current speech segment. Shorter values yield lower latency but can prematurely cut sentences when there is a brief mid-sentence pause. Longer values tolerate natural pauses better but add latency before the STT begins.
speech-to-speech --min_silence_ms 100
​
min_speech_ms
integer
default:"384"
Minimum duration of detected speech (in milliseconds) required before a segment is forwarded downstream. Frames of activity shorter than this threshold are discarded as noise. The default 384 ms is tuned to eliminate keyboard clicks and brief background sounds while accepting normal speech.
speech-to-speech --min_speech_ms 384
​
min_speech_continuation_ms
integer
default:"192"
Hysteresis threshold for speech that continues a soft-ended, uncommitted turn that is still within the speculative reopen window. When the user resumes speaking after a soft-ended turn, this lower bar (rather than min_speech_ms) is used to accept the continuation fragment. The recommended pairing is --min_speech_ms 384 --min_speech_continuation_ms 192. Set to 0 to disable the split and always require min_speech_ms. Clamped internally to [100, min_speech_ms]. Barge-in detection is unaffected.
speech-to-speech --min_speech_ms 384 --min_speech_continuation_ms 192
​
max_speech_ms
float
default:"inf"
Maximum length of a continuous speech segment in milliseconds before the VAD forces a split and forwards the accumulated audio downstream. The default is infinite, meaning very long utterances are never force-split. Set a finite value if you want to process long monologues in rolling chunks.
speech-to-speech --max_speech_ms 10000
​
speech_pad_ms
integer
default:"500"
Amount of audio (in milliseconds) retained in a ring buffer before VAD triggers and prepended to detected speech segments. This ensures that the leading edge of an utterance — which may have been partially buffered before the VAD threshold was crossed — is included in the audio sent to STT.
speech-to-speech --speech_pad_ms 300
​
audio_enhancement
boolean
default:"false"
When true, applies DeepFilterNet noise reduction, equalization, and echo cancellation to the audio before VAD processing. Can improve accuracy in noisy environments. Requires numpy<2 and is incompatible with Pocket TTS, which requires numpy>=2.
speech-to-speech --audio_enhancement
audio_enhancement requires DeepFilterNet, which conflicts with Pocket TTS (--tts pocket). Do not combine both in the same environment.
​
enable_realtime_transcription
boolean
default:"false"
When true, the VAD releases progressive audio chunks to the STT handler while the user is still speaking, enabling incremental transcription updates. Used internally by the live transcription feature; normally set via --enable_live_transcription on ModuleArguments.
speech-to-speech --enable_realtime_transcription
​
realtime_processing_pause
float
default:"0.5"
Interval in seconds between progressive audio chunk releases during active speech when enable_realtime_transcription is on. Lower values produce more frequent partial transcription updates.
speech-to-speech --enable_realtime_transcription --realtime_processing_pause 0.25
​
speculative_reopen_ms
integer
default:"1000"
In realtime mode, the number of milliseconds after a soft turn-end during which the turn stays reopenable. If the user starts speaking again within this window and no assistant output has been committed yet, the new speech continues the same turn instead of starting a new one.
speech-to-speech --speculative_reopen_ms 1500
​
unanswered_reopen_ms
integer
default:"7000"
A sanity cap (in milliseconds) for how long a soft-ended speculative turn that has not yet received any assistant output remains reopenable. This bounds the unanswered case and is not a primary tuning knob; it has no effect when speculative_reopen_ms has already expired.
speech-to-speech --unanswered_reopen_ms 5000
​
short_segment_merge_ms
integer
default:"0"
When greater than 0, adjacent VAD segments that are each shorter than min_speech_ms are held and stitched together for up to this many milliseconds before being discarded. Fragments shorter than 100 ms of active speech are never held. Useful when min_silence_ms is very low (e.g. 32 ms) to reduce the risk of splitting a single utterance into many tiny fragments.
speech-to-speech --min_silence_ms 32 --short_segment_merge_ms 300

​
Tuning for sensitivity

# Lower thresh, larger pad, audio enhancement on
speech-to-speech \
    --thresh 0.4 \
    --speech_pad_ms 700 \
    --min_silence_ms 128 \
    --audio_enhancement
The recommended baseline for most use cases is --thresh 0.6 --min_speech_ms 384 --min_speech_continuation_ms 192, which is what the default speech-to-speech command uses.

Build docs developers (and LLMs) love

Get started for freeTalk to us