• Docs >
  • Online ASR with Emformer RNN-T >
  • Nightly (unstable)
Shortcuts

Online ASR with Emformer RNN-T

Author: Jeff Hwang, Moto Hira

This tutorial shows how to use Emformer RNN-T and streaming API to perform online speech recognition.

Note

This tutorial requires FFmpeg libraries and SentencePiece.

Please refer to Optional Dependencies for the detail.

1. Overview

Performing online speech recognition is composed of the following steps

  1. Build the inference pipeline Emformer RNN-T is composed of three components: feature extractor, decoder and token processor.

  2. Format the waveform into chunks of expected sizes.

  3. Pass data through the pipeline.

2. Preparation

import torch
import torchaudio

print(torch.__version__)
print(torchaudio.__version__)

import IPython
import matplotlib.pyplot as plt
from torchaudio.io import StreamReader
2.5.0.dev20240627
2.4.0.dev20240628

3. Construct the pipeline

Pre-trained model weights and related pipeline components are bundled as torchaudio.pipelines.RNNTBundle.

We use torchaudio.pipelines.EMFORMER_RNNT_BASE_LIBRISPEECH, which is a Emformer RNN-T model trained on LibriSpeech dataset.

bundle = torchaudio.pipelines.EMFORMER_RNNT_BASE_LIBRISPEECH

feature_extractor = bundle.get_streaming_feature_extractor()
decoder = bundle.get_decoder()
token_processor = bundle.get_token_processor()
  0%|          | 0.00/3.81k [00:00<?, ?B/s]
100%|##########| 3.81k/3.81k [00:00<00:00, 4.11MB/s]

  0%|          | 0.00/293M [00:00<?, ?B/s]
 11%|#         | 31.8M/293M [00:00<00:00, 333MB/s]
 22%|##1       | 63.5M/293M [00:00<00:00, 303MB/s]
 32%|###2      | 95.1M/293M [00:00<00:00, 315MB/s]
 44%|####4     | 130M/293M [00:00<00:00, 332MB/s]
 56%|#####5    | 164M/293M [00:00<00:00, 342MB/s]
 67%|######7   | 196M/293M [00:00<00:00, 321MB/s]
 80%|########  | 235M/293M [00:00<00:00, 347MB/s]
 92%|#########1| 269M/293M [00:00<00:00, 326MB/s]
100%|##########| 293M/293M [00:00<00:00, 326MB/s]
/pytorch/audio/src/torchaudio/pipelines/rnnt_pipeline.py:248: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
  state_dict = torch.load(path)

  0%|          | 0.00/295k [00:00<?, ?B/s]
100%|##########| 295k/295k [00:00<00:00, 93.5MB/s]

Streaming inference works on input data with overlap. Emformer RNN-T model treats the newest portion of the input data as the “right context” — a preview of future context. In each inference call, the model expects the main segment to start from this right context from the previous inference call. The following figure illustrates this.

https://download.pytorch.org/torchaudio/tutorial-assets/emformer_rnnt_context.png

The size of main segment and right context, along with the expected sample rate can be retrieved from 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)")
Sample rate: 16000
Main segment: 2560 frames (0.16 seconds)
Right context: 640 frames (0.04 seconds)

4. Configure the audio stream

Next, we configure the input audio stream using torchaudio.io.StreamReader.

For the detail of this API, please refer to the StreamReader Basic Usage.

The following audio file was originally published by LibriVox project, and it is in the public domain.

https://librivox.org/great-pirate-stories-by-joseph-lewis-french/

It was re-uploaded for the sake of the tutorial.

src = "https://download.pytorch.org/torchaudio/tutorial-assets/greatpiratestories_00_various.mp3"

streamer = StreamReader(src)
streamer.add_basic_audio_stream(frames_per_chunk=segment_length, sample_rate=bundle.sample_rate)

print(streamer.get_src_stream_info(0))
print(streamer.get_out_stream_info(0))
SourceAudioStream(media_type='audio', codec='mp3', codec_long_name='MP3 (MPEG audio layer 3)', format='fltp', bit_rate=128000, num_frames=0, bits_per_sample=0, metadata={}, sample_rate=44100.0, num_channels=2)
OutputAudioStream(source_index=0, filter_description='aresample=16000,aformat=sample_fmts=fltp', media_type='audio', format='fltp', sample_rate=16000.0, num_channels=2)

As previously explained, Emformer RNN-T model expects input data with overlaps; however, Streamer iterates the source media without overlap, so we make a helper structure that caches a part of input data from Streamer as right context and then appends it to the next input data from Streamer.

The following figure illustrates this.

https://download.pytorch.org/torchaudio/tutorial-assets/emformer_rnnt_streamer_context.png
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. Run stream inference

Finally, we run the recognition.

First, we initialize the stream iterator, context cacher, and state and hypothesis that are used by decoder to carry over the decoding state between inference calls.

cacher = ContextCacher(segment_length, context_length)

state, hypothesis = None, None

Next we, run the inference.

For the sake of better display, we create a helper function which processes the source stream up to the given times and call it repeatedly.

stream_iterator = streamer.stream()


def _plot(feats, num_iter, unit=25):
    unit_dur = segment_length / sample_rate * unit
    num_plots = num_iter // unit + (1 if num_iter % unit else 0)
    fig, axes = plt.subplots(num_plots, 1)
    t0 = 0
    for i, ax in enumerate(axes):
        feats_ = feats[i * unit : (i + 1) * unit]
        t1 = t0 + segment_length / sample_rate * len(feats_)
        feats_ = torch.cat([f[2:-2] for f in feats_])  # remove boundary effect and overlap
        ax.imshow(feats_.T, extent=[t0, t1, 0, 1], aspect="auto", origin="lower")
        ax.tick_params(which="both", left=False, labelleft=False)
        ax.set_xlim(t0, t0 + unit_dur)
        t0 = t1
    fig.suptitle("MelSpectrogram Feature")
    plt.tight_layout()


@torch.inference_mode()
def run_inference(num_iter=100):
    global state, hypothesis
    chunks = []
    feats = []
    for i, (chunk,) in enumerate(stream_iterator, start=1):
        segment = cacher(chunk[:, 0])
        features, length = feature_extractor(segment)
        hypos, state = decoder.infer(features, length, 10, state=state, hypothesis=hypothesis)
        hypothesis = hypos
        transcript = token_processor(hypos[0][0], lstrip=False)
        print(transcript, end="\r", flush=True)

        chunks.append(chunk)
        feats.append(features)
        if i == num_iter:
            break

    # Plot the features
    _plot(feats, num_iter)
    return IPython.display.Audio(torch.cat(chunks).T.numpy(), rate=bundle.sample_rate)
run_inference()
MelSpectrogram Feature
forward
forward
forward
forward
forward
forward great
forward great pir
forward great pirate
forward great pirate
forward great pirate stories
forward great pirate stories
forward great pirate stories
forward great pirate stories
forward great pirate stories
forward great pirate stories
forward great pirate stories
forward great pirate stories this
forward great pirate stories this is
forward great pirate stories this is a
forward great pirate stories this is a liber
forward great pirate stories this is a liberal
forward great pirate stories this is a liberal
forward great pirate stories this is a liberal record
forward great pirate stories this is a liberal record
forward great pirate stories this is a liberal recording
forward great pirate stories this is a liberal recording
forward great pirate stories this is a liberal recording all
forward great pirate stories this is a liberal recording all the
forward great pirate stories this is a liberal recording all the
forward great pirate stories this is a liberal recording all the votes
forward great pirate stories this is a liberal recording all the votes record
forward great pirate stories this is a liberal recording all the votes record
forward great pirate stories this is a liberal recording all the votes recordings
forward great pirate stories this is a liberal recording all the votes recordings are
forward great pirate stories this is a liberal recording all the votes recordings are in the
forward great pirate stories this is a liberal recording all the votes recordings are in the public
forward great pirate stories this is a liberal recording all the votes recordings are in the public
forward great pirate stories this is a liberal recording all the votes recordings are in the public dum
forward great pirate stories this is a liberal recording all the votes recordings are in the public dum
forward great pirate stories this is a liberal recording all the votes recordings are in the public duma
forward great pirate stories this is a liberal recording all the votes recordings are in the public duma
forward great pirate stories this is a liberal recording all the votes recordings are in the public duma
forward great pirate stories this is a liberal recording all the votes recordings are in the public duma for more
forward great pirate stories this is a liberal recording all the votes recordings are in the public duma for more
forward great pirate stories this is a liberal recording all the votes recordings are in the public duma for more information
forward great pirate stories this is a liberal recording all the votes recordings are in the public duma for more information
forward great pirate stories this is a liberal recording all the votes recordings are in the public duma for more information
forward great pirate stories this is a liberal recording all the votes recordings are in the public duma for more information or
forward great pirate stories this is a liberal recording all the votes recordings are in the public duma for more information or develop
forward great pirate stories this is a liberal recording all the votes recordings are in the public duma for more information or to volunt
forward great pirate stories this is a liberal recording all the votes recordings are in the public duma for more information or to volunteer
forward great pirate stories this is a liberal recording all the votes recordings are in the public duma for more information or to volunteer
forward great pirate stories this is a liberal recording all the votes recordings are in the public duma for more information or to volunteer
forward great pirate stories this is a liberal recording all the votes recordings are in the public duma for more information or to volunteer please
forward great pirate stories this is a liberal recording all the votes recordings are in the public duma for more information or to volunteer please
forward great pirate stories this is a liberal recording all the votes recordings are in the public duma for more information or to volunteer please visit
forward great pirate stories this is a liberal recording all the votes recordings are in the public duma for more information or to volunteer please visit
forward great pirate stories this is a liberal recording all the votes recordings are in the public duma for more information or to volunteer please visit liber
forward great pirate stories this is a liberal recording all the votes recordings are in the public duma for more information or to volunteer please visit liber
forward great pirate stories this is a liberal recording all the votes recordings are in the public duma for more information or to volunteer please visit liberal
forward great pirate stories this is a liberal recording all the votes recordings are in the public dummy for more information or to volunteer please visit liberal
forward great pirate stories this is a liberal recording all the votes recordings are in the public dummy for more information or to volunteer please visit liberal stout
forward great pirate stories this is a liberal recording all the votes recordings are in the public dummy for more information or to volunteer please visit liberal stout work
forward great pirate stories this is a liberal recording all the votes recordings are in the public dummy for more information or to volunteer please visit liberal stout work
forward great pirate stories this is a liberal recording all the votes recordings are in the public dummy for more information or to volunteer please visit liberal stout work
forward great pirate stories this is a liberal recording all the votes recordings are in the public dummy for more information or to volunteer please visit liberal stout work
forward great pirate stories this is a liberal recording all the votes recordings are in the public dummy for more information or to volunteer please visit liberal stout work
forward great pirate stories this is a liberal recording all the votes recordings are in the public dummy for more information or to volunteer please visit liberal stout work
forward great pirate stories this is a liberal recording all the votes recordings are in the public dummy for more information or to volunteer please visit liberal stout work
forward great pirate stories this is a liberal recording all the votes recordings are in the public dummy for more information or to volunteer please visit liberal stout work
forward great pirate stories this is a liberal recording all the votes recordings are in the public dummy for more information or to volunteer please visit liberal stout work record
forward great pirate stories this is a liberal recording all the votes recordings are in the public dummy for more information or to volunteer please visit liberal stout work recording
forward great pirate stories this is a liberal recording all the votes recordings are in the public dummy for more information or to volunteer please visit liberal stout work recording by
forward great pirate stories this is a liberal recording all the votes recordings are in the public dummy for more information or to volunteer please visit liberal stout work recording by james
forward great pirate stories this is a liberal recording all the votes recordings are in the public dummy for more information or to volunteer please visit liberal stout work recording by james
forward great pirate stories this is a liberal recording all the votes recordings are in the public dummy for more information or to volunteer please visit liberal stout work recording by james
forward great pirate stories this is a liberal recording all the votes recordings are in the public dummy for more information or to volunteer please visit liberal stout work recording by james christ
forward great pirate stories this is a liberal recording all the votes recordings are in the public dummy for more information or to volunteer please visit liberal stout work recording by james christopher
forward great pirate stories this is a liberal recording all the votes recordings are in the public dummy for more information or to volunteer please visit liberal stout work recording by james christopher
forward great pirate stories this is a liberal recording all the votes recordings are in the public dummy for more information or to volunteer please visit liberal stout work recording by james christopher
forward great pirate stories this is a liberal recording all the votes recordings are in the public dummy for more information or to volunteer please visit liberal stout work recording by james christopher
forward great pirate stories this is a liberal recording all the votes recordings are in the public dummy for more information or to volunteer please visit liberal stout work recording by james christopher
forward great pirate stories this is a liberal recording all the votes recordings are in the public dummy for more information or to volunteer please visit liberal stout work recording by james christopher
forward great pirate stories this is a liberal recording all the votes recordings are in the public dummy for more information or to volunteer please visit liberal stout work recording by james christopher
forward great pirate stories this is a liberal recording all the votes recordings are in the public dummy for more information or to volunteer please visit liberal stout work recording by james christopher great
forward great pirate stories this is a liberal recording all the votes recordings are in the public dummy for more information or to volunteer please visit liberal stout work recording by james christopher great
forward great pirate stories this is a liberal recording all the votes recordings are in the public dummy for more information or to volunteer please visit liberal stout work recording by james christopher great pir
forward great pirate stories this is a liberal recording all the votes recordings are in the public dummy for more information or to volunteer please visit liberal stout work recording by james christopher great pirate
forward great pirate stories this is a liberal recording all the votes recordings are in the public dummy for more information or to volunteer please visit liberal stout work recording by james christopher great pirate
forward great pirate stories this is a liberal recording all the votes recordings are in the public dummy for more information or to volunteer please visit liberal stout work recording by james christopher great pirate stories
forward great pirate stories this is a liberal recording all the votes recordings are in the public dummy for more information or to volunteer please visit liberal stout work recording by james christopher great pirate stories
forward great pirate stories this is a liberal recording all the votes recordings are in the public dummy for more information or to volunteer please visit liberal stout work recording by james christopher great pirate stories
forward great pirate stories this is a liberal recording all the votes recordings are in the public dummy for more information or to volunteer please visit liberal stout work recording by james christopher great pirate stories