[BUG] last audio chunk is sent to processor twice.

[lilithgrigoryan]

Hi everyone,

While looking at the code in inference.py, I noticed that the last audio chunk appears to be sent to the inference engine twice: once inside the loop and then again in the following if block.

simulstream/simulstream/inference.py

Lines 40 to 63 in 8a14aa1

	def process_audio(
	message_processor: MessageProcessor,
	sample_rate: int,
	data: np.ndarray):
	"""
	Stream audio data in fixed-size chunks over a WebSocket connection.
	Args:
	message_processor (MessageProcessor): class that processes the audio chunks.
	sample_rate (int): Audio sample rate (Hz).
	data (np.ndarray): Audio samples as int16 array.
	"""
	# speech_chunk_size is expressed in seconds, so the number of samples corresponding to
	# one speech chunk is the following
	samples_per_chunk = int(
	sample_rate * message_processor.speech_processor.speech_chunk_size)
	i = 0
	for i in range(0, len(data), samples_per_chunk):
	output = message_processor.process_speech(data[i:i + samples_per_chunk].tobytes())
	LOGGER.debug(f"response: {output}")
	# send last part of the audio
	if i < len(data):
	output = message_processor.process_speech(data[i:].tobytes())
	LOGGER.debug(f"response: {output}")

It seems that simplifying the loop slightly might avoid this behavior. For example:

def process_audio(
    message_processor: MessageProcessor,
    sample_rate: int,
    data: np.ndarray,
):
    """
    Stream audio data in fixed-size chunks over a WebSocket connection.
    """
    samples_per_chunk = int(
        sample_rate * message_processor.speech_processor.speech_chunk_size
    )

    for start in range(0, len(data), samples_per_chunk):
        chunk = data[start:start + samples_per_chunk]

        if len(chunk) == 0:
            continue

        output = message_processor.process_speech(chunk.tobytes())
        LOGGER.debug(f"response: {output}")

Could you please take a look, and if possible fix this.

[sarapapi]

Thanks for spotting this issue. @mgaido91 I think we can avoid that if as well, if you agree, I'll take care of this

[mgaido91]

@lilithgrigoryan thanks for raising the issue. We need to fix this. I am assigning this to you then @sarapapi