How to Adjust chunk_size for 1-Second Audio Input in Cache-Aware Streaming ASR Demo?

[june-oh]

Hi, I’m using the Online_ASR_Microphone_Demo_Cache_Aware_Streaming.ipynb notebook and referring to the following code snippet:

with torch.no_grad():
    (
        pred_out_stream,
        transcribed_texts,
        cache_last_channel,
        cache_last_time,
        cache_last_channel_len,
        previous_hypotheses,
    ) = asr_model.conformer_stream_step(
        processed_signal=processed_signal,
        processed_signal_length=processed_signal_length,
        cache_last_channel=cache_last_channel,
        cache_last_time=cache_last_time,
        cache_last_channel_len=cache_last_channel_len,
        keep_all_outputs=False,
        previous_hypotheses=previous_hypotheses,
        previous_pred_out=pred_out_stream,
        drop_extra_pre_encoded=None,
        return_transcription=True,
    )

In this demo, the chunk_size is defined as chunk_size = lookahead_size + ENCODER_STEP_LENGTH, but the default setting results in chunks that are too small for my use case. I’d like to process the audio input in approximately 1-second increments.

However, simply changing ENCODER_STEP_LENGTH to 1000 didn’t produce the expected results. Could anyone provide guidance on how to properly adjust the chunk_size or any other relevant parameters to achieve this? Any detailed explanation or example would be greatly appreciated!