Uses an audio-classification model to verify that audio flagged as speech by an upstream VAD node is genuine speech, suppressing speech segments that turn out to be non-speech noise before they reach the rest of the flow.
Parameters
name |
description | default |
|---|---|---|
| Minimum Bypass Duration(ms) | The minimum duration for bypassing the filter in milliseconds. Once a speech segment's buffered audio exceeds this duration without having ended yet, the node doesn't check the segment with classification and lets the segment through as speech. | 1500 |
| Sensitivity | Sensitivity between 0.0 and 1.0. This is the minimum aggregated speech-class score a segment must reach to be classified as speech. | 0.5 |
Inputs
Audio
While a speech segment is being collected, incoming audio is buffered internally rather than forwarded. Once a segment has been accepted as speech, any further audio for that segment is passed straight through without buffering.
Events
| name | description | possible nodes this event may be received from |
|---|---|---|
| Speech Started | Clears any previous state and starts buffering audio for a new speech segment. | VAD Elastic, Silero Vad, Vad Service |
| Early Speech Ended | If the segment has already been accepted as speech, it is forwarded immediately. Otherwise, triggers YAMNet inference on the audio buffered so far; the buffer is forwarded if the result is "speech". | VAD Elastic |
| Speech Ended | If the segment has already been accepted as speech, forwarded immediately and the node's state is cleared. Otherwise, triggers YAMNet inference on the buffered audio. Forwards buffer if result is "Speech". | VAD Elastic, Silero Vad, Vad Service |
Outputs
Audio
The audio buffered for a speech segment is forwarded once the segment is accepted as speech — either because the model classified it as speech, or because Minimum Bypass Duration was exceeded before a result could be produced. Audio for segments classified as non-speech is discarded and never forwarded.
Events
| name | description |
|---|---|
| Speech Started / Speech Ended / Early Speech Ended | Forwards incoming events as they were, only if result is "Speech" |
| Noise Filter Result | Reports the outcome of a filtering decision for a segment. |
Remarks
Project Structure
The node outputs mirror any VAD node outputs, so the downstream can look exactly like the output of VAD.
Supported flow types
Stream