Device ASR with Emformer RNN-T

Author: Moto Hira, Jeff Hwang.

This tutorial shows how to use Emformer RNN-T and streaming API to perform speech recognition on a streaming device input, i.e. microphone on laptop.

Note

This tutorial requires FFmpeg libraries (>=4.1, <4.4) and SentencePiece.

There are multiple ways to install FFmpeg libraries. If you are using Anaconda Python distribution, conda install 'ffmpeg<4.4' will install the required FFmpeg libraries.

You can install SentencePiece by running pip install sentencepiece.

Note

This tutorial was tested on MacBook Pro and Dynabook with Windows 10.

This tutorial does NOT work on Google Colab because the server running this tutorial does not have a microphone that you can talk to.

1. Overview

We use streaming API to fetch audio from audio device (microphone) chunk by chunk, then run inference using Emformer RNN-T.

For the basic usage of the streaming API and Emformer RNN-T please refer to StreamReader Basic Usage and Online ASR with Emformer RNN-T.

2. Checking the supported devices

Firstly, we need to check the devices that Streaming API can access, and figure out the arguments (src and format) we need to pass to StreamReader() class.

We use ffmpeg command for this. ffmpeg abstracts away the difference of underlying hardware implementations, but the expected value for format varies across OS and each format defines different syntax for src.

The details of supported format values and src syntax can be found in https://ffmpeg.org/ffmpeg-devices.html.

For macOS, the following command will list the available devices.

$ ffmpeg -f avfoundation -list_devices true -i dummy
...
[AVFoundation indev @ 0x126e049d0] AVFoundation video devices:
[AVFoundation indev @ 0x126e049d0] [0] FaceTime HD Camera
[AVFoundation indev @ 0x126e049d0] [1] Capture screen 0
[AVFoundation indev @ 0x126e049d0] AVFoundation audio devices:
[AVFoundation indev @ 0x126e049d0] [0] ZoomAudioDevice
[AVFoundation indev @ 0x126e049d0] [1] MacBook Pro Microphone

We will use the following values for Streaming API.

StreamReader(
    src = ":1",  # no video, audio from device 1, "MacBook Pro Microphone"
    format = "avfoundation",
)

For Windows, dshow device should work.

> ffmpeg -f dshow -list_devices true -i dummy
...
[dshow @ 000001adcabb02c0] DirectShow video devices (some may be both video and audio devices)
[dshow @ 000001adcabb02c0]  "TOSHIBA Web Camera - FHD"
[dshow @ 000001adcabb02c0]     Alternative name "@device_pnp_\\?\usb#vid_10f1&pid_1a42&mi_00#7&27d916e6&0&0000#{65e8773d-8f56-11d0-a3b9-00a0c9223196}\global"
[dshow @ 000001adcabb02c0] DirectShow audio devices
[dshow @ 000001adcabb02c0]  "... (Realtek High Definition Audio)"
[dshow @ 000001adcabb02c0]     Alternative name "@device_cm_{33D9A762-90C8-11D0-BD43-00A0C911CE86}\wave_{BF2B8AE1-10B8-4CA4-A0DC-D02E18A56177}"

In the above case, the following value can be used to stream from microphone.

StreamReader(
    src = "audio=@device_cm_{33D9A762-90C8-11D0-BD43-00A0C911CE86}\wave_{BF2B8AE1-10B8-4CA4-A0DC-D02E18A56177}",
    format = "dshow",
)

3. Data acquisition

Streaming audio from microphone input requires properly timing data acquisition. Failing to do so may introduce discontinuities in the data stream.

For this reason, we will run the data acquisition in a subprocess.

Firstly, we create a helper function that encapsulates the whole process executed in the subprocess.

This function initializes the streaming API, acquires data then puts it in a queue, which the main process is watching.

import torch
import torchaudio


# The data acquisition process will stop after this number of steps.
# This eliminates the need of process synchronization and makes this
# tutorial simple.
NUM_ITER = 100


def stream(q, format, src, segment_length, sample_rate):
    from torchaudio.io import StreamReader

    print("Building StreamReader...")
    streamer = StreamReader(src, format=format)
    streamer.add_basic_audio_stream(frames_per_chunk=segment_length, sample_rate=sample_rate)

    print(streamer.get_src_stream_info(0))
    print(streamer.get_out_stream_info(0))

    print("Streaming...")
    print()
    stream_iterator = streamer.stream(timeout=-1, backoff=1.0)
    for _ in range(NUM_ITER):
        (chunk,) = next(stream_iterator)
        q.put(chunk)

The notable difference from the non-device streaming is that, we provide timeout and backoff parameters to stream method.

When acquiring data, if the rate of acquisition requests is higher than that at which the hardware can prepare the data, then the underlying implementation reports special error code, and expects client code to retry.

Precise timing is the key for smooth streaming. Reporting this error from low-level implementation all the way back to Python layer, before retrying adds undesired overhead. For this reason, the retry behavior is implemented in C++ layer, and timeout and backoff parameters allow client code to control the behavior.

For the detail of timeout and backoff parameters, please refer to the documentation of stream() method.

Note

The proper value of backoff depends on the system configuration. One way to see if backoff value is appropriate is to save the series of acquired chunks as a continuous audio and listen to it. If backoff value is too large, then the data stream is discontinuous. The resulting audio sounds sped up. If backoff value is too small or zero, the audio stream is fine, but the data acquisition process enters busy-waiting state, and this increases the CPU consumption.

4. Building inference pipeline

The next step is to create components required for inference.

This is the same process as Online ASR with Emformer RNN-T.

class Pipeline:
    """Build inference pipeline from RNNTBundle.

    Args:
        bundle (torchaudio.pipelines.RNNTBundle): Bundle object
        beam_width (int): Beam size of beam search decoder.
    """

    def __init__(self, bundle: torchaudio.pipelines.RNNTBundle, beam_width: int = 10):
        self.bundle = bundle
        self.feature_extractor = bundle.get_streaming_feature_extractor()
        self.decoder = bundle.get_decoder()
        self.token_processor = bundle.get_token_processor()

        self.beam_width = beam_width

        self.state = None
        self.hypothesis = None

    def infer(self, segment: torch.Tensor) -> str:
        """Perform streaming inference"""
        features, length = self.feature_extractor(segment)
        hypos, self.state = self.decoder.infer(
            features, length, self.beam_width, state=self.state, hypothesis=self.hypothesis
        )
        self.hypothesis = hypos[0]
        transcript = self.token_processor(self.hypothesis[0], lstrip=False)
        return transcript

class ContextCacher:
    """Cache the end of input data and prepend the next input data with it.

    Args:
        segment_length (int): The size of main segment.
            If the incoming segment is shorter, then the segment is padded.
        context_length (int): The size of the context, cached and appended.
    """

    def __init__(self, segment_length: int, context_length: int):
        self.segment_length = segment_length
        self.context_length = context_length
        self.context = torch.zeros([context_length])

    def __call__(self, chunk: torch.Tensor):
        if chunk.size(0) < self.segment_length:
            chunk = torch.nn.functional.pad(chunk, (0, self.segment_length - chunk.size(0)))
        chunk_with_context = torch.cat((self.context, chunk))
        self.context = chunk[-self.context_length :]
        return chunk_with_context

5. The main process

The execution flow of the main process is as follows:

1. Initialize the inference pipeline.

2. Launch data acquisition subprocess.

3. Run inference.

4. Clean up

Note

As the data acquisition subprocess will be launched with “spawn” method, all the code on global scope are executed on the subprocess as well.

We want to instantiate pipeline only in the main process, so we put them in a function and invoke it within __name__ == “__main__” guard.

def main(device, src, bundle):
    print(torch.__version__)
    print(torchaudio.__version__)

    print("Building pipeline...")
    pipeline = Pipeline(bundle)

    sample_rate = bundle.sample_rate
    segment_length = bundle.segment_length * bundle.hop_length
    context_length = bundle.right_context_length * bundle.hop_length

    print(f"Sample rate: {sample_rate}")
    print(f"Main segment: {segment_length} frames ({segment_length / sample_rate} seconds)")
    print(f"Right context: {context_length} frames ({context_length / sample_rate} seconds)")

    cacher = ContextCacher(segment_length, context_length)

    @torch.inference_mode()
    def infer():
        for _ in range(NUM_ITER):
            chunk = q.get()
            segment = cacher(chunk[:, 0])
            transcript = pipeline.infer(segment)
            print(transcript, end="", flush=True)

    import torch.multiprocessing as mp

    ctx = mp.get_context("spawn")
    q = ctx.Queue()
    p = ctx.Process(target=stream, args=(q, device, src, segment_length, sample_rate))
    p.start()
    infer()
    p.join()


if __name__ == "__main__":
    main(
        device="avfoundation",
        src=":1",
        bundle=torchaudio.pipelines.EMFORMER_RNNT_BASE_LIBRISPEECH,
    )

Building pipeline...
Sample rate: 16000
Main segment: 2560 frames (0.16 seconds)
Right context: 640 frames (0.04 seconds)
Building StreamReader...
SourceAudioStream(media_type='audio', codec='pcm_f32le', codec_long_name='PCM 32-bit floating point little-endian', format='flt', bit_rate=1536000, sample_rate=48000.0, num_channels=1)
OutputStream(source_index=0, filter_description='aresample=16000,aformat=sample_fmts=fltp')
Streaming...

hello world

Tag: torchaudio.io