metadata

language: en
thumbnail: null
tags:
  - ASR
  - CTC
  - Attention
  - pytorch
license: apache-2.0
datasets:
  - librispeech
metrics:
  - wer
  - cer

CRDNN with CTC/Attention and RNNLM trained on LibriSpeech

This repository provides all the necessary tools to perform automatic speech recognition from an end-to-end system pretrained on LibriSpeech (EN) within SpeechBrain. For a better experience we encourage you to learn more about SpeechBrain. The given ASR model performance are:

Release	hyperparams file	Test WER	Model link	GPUs
20-05-22	BPE_1000.yaml	3.08	Not Available	1xV100 32GB
20-05-22	BPE_5000.yaml	2.89	Not Available	1xV100 32GB

Pipeline description

This ASR system is composed with 3 different but linked blocks:

Tokenizer (unigram) that transforms words into subword units and trained with the train transcriptions of LibriSpeech.
Neural language model (RNNLM) trained on the full 10M words dataset.
Acoustic model (CRDNN + CTC/Attention). The CRDNN architecture is made of N blocks of convolutional neural networks with normalisation and pooling on the frequency domain. Then, a bidirectional LSTM is connected to a final DNN to obtain the final acoustic representation that is given to the CTC and attention decoders.

Intended uses & limitations

This model has been primilarly developed to be run within SpeechBrain as a pretrained ASR model for the english language. Thanks to the flexibility of SpeechBrain, any of the 3 blocks detailed above can be extracted and connected to you custom pipeline as long as SpeechBrain is installed.

Install SpeechBrain

First of all, please install SpeechBrain with the following command:

pip install \\we hide ! SpeechBrain is still private :p

Please notice that we encourage you to read our tutorials and learn more about SpeechBrain.

Transcribing your own audio files

import torch
import torchaudio
import speechbrain
from speechbrain.lobes.pretrained.librispeech.asr_crdnn_ctc_att_rnnlm.acoustic import ASR

asr_model = ASR()

# Make sure your output is sampled at 16 kHz.
audio_file='path_to_your_audio_file'
wav, fs = torchaudio.load(audio_file)
wav_lens = torch.tensor([1]).float()

# Transcribe!
words, tokens = asr_model.transcribe(wav, wav_lens)
print(words)

Obtaining encoded features

The SpeechBrain ASR() Class provides an easy way to encode the speech signal without running the decoding phase. Hence, one can obtain the output of the CRDNN model.

import torch
import torchaudio
import speechbrain
from speechbrain.lobes.pretrained.librispeech.asr_crdnn_ctc_att_rnnlm.acoustic import ASR

asr_model = ASR()

# Make sure your output is sampled at 16 kHz.
audio_file='path_to_your_audio_file'
wav, fs = torchaudio.load(audio_file)
wav_lens = torch.tensor([1]).float()

# Transcribe!
words, tokens = asr_model.encode(wav, wav_lens)
print(words)

Playing with the language model only

Thanks to SpeechBrain lobes, it is feasible to simply instantiate the language model to further processing on your custom pipeline:

import torch
import speechbrain
from speechbrain.lobes.pretrained.librispeech.asr_crdnn_ctc_att_rnnlm.lm import LM

lm = LM()

text = "THE CAT IS ON"

# Next word prediction
encoded_text = lm.tokenizer.encode_as_ids(text)
encoded_text = torch.Tensor(encoded_text).unsqueeze(0)
prob_out, _ = lm(encoded_text.to(lm.device))
index = int(torch.argmax(prob_out[0,-1,:]))
print(lm.tokenizer.decode(index))

# Text generation
encoded_text = torch.tensor([0, 2]) # bos token + the
encoded_text = encoded_text.unsqueeze(0).to(lm.device)
for i in range(19):
  prob_out, _ = lm(encoded_text)
  index = torch.argmax(prob_out[0,-1,:]).unsqueeze(0)
  encoded_text = torch.cat([encoded_text, index.unsqueeze(0)], dim=1)
encoded_text = encoded_text[0,1:].tolist()
print(lm.tokenizer.decode(encoded_text))

Playing with the tokenizer only

In the same manner as for the language model, one can isntantiate the tokenizer only with the corresponding lobes in SpeechBrain.

import speechbrain
from speechbrain.lobes.pretrained.librispeech.asr_crdnn_ctc_att_rnnlm.tokenizer import tokenizer

# HuggingFace paths to download the pretrained models
token_file = 'tokenizer/1000_unigram.model'
model_name = 'sb/asr-crdnn-librispeech'
save_dir = 'model_checkpoints'

text = "THE CAT IS ON THE TABLE"

tokenizer = tokenizer(token_file, model_name, save_dir)

# Tokenize!
print(tokenizer.spm.encode(text))
print(tokenizer.spm.encode(text, out_type='str'))

Referencing SpeechBrain

@misc{SB2021,
    author = {Ravanelli, Mirco and Parcollet, Titouan and Rouhe, Aku and Plantinga, Peter and Rastorgueva, Elena and Lugosch, Loren and Dawalatabad, Nauman and Ju-Chieh, Chou and Heba, Abdel and Grondin, Francois and Aris, William and Liao, Chien-Feng and Cornell, Samuele and Yeh, Sung-Lin and Na, Hwidong and Gao, Yan and Fu, Szu-Wei and Subakan, Cem and De Mori, Renato and Bengio, Yoshua },
    title = {SpeechBrain},
    year = {2021},
    publisher = {GitHub},
    journal = {GitHub repository},
    howpublished = {\url{https://github.com/speechbrain/speechbrain}},
  }