Silero adapter cuts off prefixPaddingSamples in non-streaming STT

[ianjaku]

Describe the bug

The Silero adapter cuts off the prefixPaddingSamples before calling _recognize when not streaming.

This causes the following:

1. We pass prefixPaddingDuration as 500ms so that we receive the audio right before the VAD fires
2. Right before _recognize gets called in vad.ts, the adapter cuts off the prefix.

The culprit is line 306 on 77a8355: this.#speechBuffer.subarray(this.#prefixPaddingSamples, speechBufferIndex)
Which should be this.#speechBuffer.subarray(0, speechBufferIndex) instead.

Relevant log output

No response

Describe your environment

We're running agents framework 1.2.7 and using the Silero adapter version 1.2.7, but we verified the same issue exists on the latest version of Silero.

Minimal reproducible example

1. Set up the most basic Livekit pipeline
2. Create a custom STT handler by extending the stt.STT class
3. Implement the _recognize class
4. Pass prefixPaddingDuration: 500 to the Silero settings
5. The audio buffer in _recognize will not receive the 500ms before the VAD fires

Additional information

No response

[ianjaku]

looking into it further, it should be a slice instead of a subarray. Otherwise resetWriteCursor overwrites it.

Makes the change:

-              this.#speechBuffer.subarray(this.#prefixPaddingSamples, speechBufferIndex),
+              this.#speechBuffer.slice(0, speechBufferIndex),