Difference Between `--min_speech_ms` and `--min_speech_continuation_ms` VAD Parameters

While --min_speech_ms defines the minimum duration required to trigger a new speech turn, --min_speech_continuation_ms acts as a lower hysteresis threshold for extending a turn that hasn't been fully committed, preventing the system from splitting continuous utterances during brief pauses.

The Hugging Face speech-to-speech repository implements a Voice Activity Detection (VAD) state machine to manage conversational turn-taking. These two VAD parameters work in tandem to balance sensitivity against false triggers, with distinct roles in the speech segmentation logic found in vad_arguments.py and vad_handler.py.

Starting New Turns with --min_speech_ms

The --min_speech_ms parameter (default 384 ms) establishes the minimum length of continuous speech required before the pipeline emits a SpeechStartedEvent and opens a new conversational turn. This threshold filters out transient noises, breath sounds, or accidental utterances that should not initiate a response cycle.

In VADHandlerArguments, this value is mapped to the min_speech_ms field. The VADHandler class retrieves this baseline threshold through the _active_speech_min_ms method when evaluating whether sufficient speech has accumulated to declare a brand-new turn. If the detected active speech duration falls below this value, the audio chunk is discarded or buffered without triggering the downstream speech-to-text pipeline.

Continuing Turns with --min_speech_continuation_ms

The --min_speech_continuation_ms parameter (default 192 ms) provides a shorter, secondary threshold specifically for turns that are "soft-ended" or pending reopen. When a user pauses briefly but hasn't finished speaking, this hysteresis value allows the system to accept a shorter burst of speech to continue the existing turn rather than treating it as a new interaction.

According to the source code in vad_handler.py, this value is resolved through _resolve_min_speech_continuation_ms, which clamps the input to the range [100, min_speech_ms]. If the parameter is set to 0 or a negative value, the method falls back to using min_speech_ms, effectively disabling the continuation relaxation.

Threshold Selection Logic in vad_handler.py

The VADHandler dynamically selects between these two thresholds via the _active_speech_min_ms method. During execution, the handler checks whether the current turn state is reopenable or pending a reopen. If so, it returns the min_speech_continuation_ms value; otherwise, it enforces the stricter min_speech_ms requirement.

This state-aware approach ensures that:

• New speech must satisfy the higher min_speech_ms bar to start a turn.
• Continuations during tentative turn closures only need to satisfy the lower min_speech_continuation_ms bar.

Configuration Examples

You can adjust these VAD parameters when launching the realtime server or instantiating the handler programmatically.

Command-line usage with CLI flags (--minspeechms and --minspeechcontinuationms):

python -m speech_to_speech.server \
    --thresh 0.7 \
    --minspeechms 500 \
    --minspeechcontinuationms 250

Programmatic configuration using VADHandlerArguments:

from speech_to_speech.arguments_classes.vad_arguments import VADHandlerArguments

args = VADHandlerArguments(
    min_speech_ms=500,
    min_speech_continuation_ms=250,
)
handler = VADHandler(**args.__dict__)

The default value of 192 ms for continuation is explicitly recommended as approximately half of the default 384 ms new-turn threshold, providing a balanced trade-off between responsiveness and turn stability.

Summary

• --min_speech_ms (default 384 ms) sets the strict minimum for detecting the start of a new speech turn, preventing false triggers from brief noises.
• --min_speech_continuation_ms (default 192 ms) provides a lower hysteresis threshold for extending turns that haven't been fully committed, improving conversational flow.
• The _resolve_min_speech_continuation_ms method enforces a valid range of [100, min_speech_ms] and falls back to min_speech_ms if the continuation value is ≤ 0.
• The _active_speech_min_ms method dynamically returns the appropriate threshold based on whether the current turn is reopenable.

Frequently Asked Questions

What happens if I set --min_speech_continuation_ms to 0?

The _resolve_min_speech_continuation_ms method in vad_handler.py treats values ≤ 0 as invalid and falls back to using min_speech_ms. This means both new turns and continuations will require the same minimum speech duration, effectively disabling the continuation hysteresis.

Why is --min_speech_continuation_ms clamped to a minimum of 100ms?

The clamping logic prevents users from setting overly aggressive continuation thresholds that could capture background noise or non-speech artifacts. By enforcing a floor of 100ms in _resolve_min_speech_continuation_ms, the system maintains a baseline level of signal quality before considering a turn extension valid.

Can --min_speech_continuation_ms be higher than --min_speech_ms?

No. The resolution method explicitly clamps the continuation value to the inclusive range [100, min_speech_ms]. If you attempt to set a higher value, it will be capped at the min_speech_ms value, ensuring that continuation logic never requires more speech than starting a fresh turn.

Where are these VAD parameters defined in the source code?

The parameters are defined in src/speech_to_speech/arguments_classes/vad_arguments.py within the VADHandlerArguments class. The runtime logic that consumes these values, including _active_speech_min_ms and _resolve_min_speech_continuation_ms, resides in src/speech_to_speech/VAD/vad_handler.py.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}