DeepAsr
DeepAsr is an open-source & Keras (Tensorflow) implementation of end-to-end Automatic Speech Recognition (ASR) engine and it supports multiple Speech Recognition architectures.
Supported Asr Architectures:
- Baidu's Deep Speech 2
- DeepAsrNetwork1
Using DeepAsr you can:
- perform speech-to-text using pre-trained models
- tune pre-trained models to your needs
- create new models on your own
DeepAsr key features:
- Multi GPU support: You can do much more like distribute the training using the Strategy, or experiment with mixed precision policy.
- CuDNN support: Model using CuDNNLSTM implementation by NVIDIA Developers. CPU devices is also supported.
- DataGenerator: The feature extraction during model training for large the data.
Installation
You can use pip:
pip install deepasr
Getting started
The speech recognition is a tough task. You don't need to know all details to use one of the pretrained models. However it's worth to understand conceptional crucial components:
- Input: Audio files (WAV or FLAC) with mono 16-bit 16 kHz (up to 5 seconds)
- FeaturesExtractor: Convert audio files using MFCC Features or Spectrogram
- Model: CTC model defined in Keras (references: [1], [2])
- Decoder: Greedy or BeamSearch algorithms with the language model support decode a sequence of probabilities using Alphabet
- DataGenerator: Stream data to the model via generator
- Callbacks: Set of functions monitoring the training
import numpy as np
import pandas as pd
import tensorflow as tf
import deepasr as asr
# get CTCPipeline
def get_config(feature_type: str = 'spectrogram', multi_gpu: bool = False):
# audio feature extractor
features_extractor = asr.features.preprocess(feature_type=feature_type, features_num=161,
samplerate=16000,
winlen=0.02,
winstep=0.025,
winfunc=np.hanning)
# input label encoder
alphabet_en = asr.vocab.Alphabet(lang='en')
# training model
model = asr.model.get_deepspeech2(
input_dim=161,
output_dim=29,
is_mixed_precision=True
)
# model optimizer
optimizer = tf.keras.optimizers.Adam(
lr=1e-4,
beta_1=0.9,
beta_2=0.999,
epsilon=1e-8
)
# output label deocder
decoder = asr.decoder.GreedyDecoder()
# decoder = asr.decoder.BeamSearchDecoder(beam_width=100, top_paths=1)
# CTCPipeline
pipeline = asr.pipeline.ctc_pipeline.CTCPipeline(
alphabet=alphabet_en, features_extractor=features_extractor, model=model, optimizer=optimizer, decoder=decoder,
sample_rate=16000, mono=True, multi_gpu=multi_gpu
)
return pipeline
train_data = pd.read_csv('train_data.csv')
pipeline = get_config(feature_type = 'fbank', multi_gpu=False)
# train asr model
history = pipeline.fit(train_dataset=train_data, batch_size=128, epochs=500)
# history = pipeline.fit_generator(train_dataset = train_data, batch_size=32, epochs=500)
pipeline.save('./checkpoint')
import pandas as pd
import tensorflow as tf
import deepasr as asr
# get CTCPipeline
def get_config(feature_type: str = 'spectrogram', multi_gpu: bool = False):
# audio feature extractor
features_extractor = asr.features.preprocess(feature_type=feature_type, features_num=161,
samplerate=16000,
winlen=0.02,
winstep=0.025,
winfunc=np.hanning)
# input label encoder
alphabet_en = asr.vocab.Alphabet(lang='en')
# training model
model = asr.model.get_deepspeech2(
input_dim=161,
output_dim=29,
is_mixed_precision=True
)
# model optimizer
optimizer = tf.keras.optimizers.Adam(
lr=1e-4,
beta_1=0.9,
beta_2=0.999,
epsilon=1e-8
)
# output label deocder
decoder = asr.decoder.GreedyDecoder()
# decoder = asr.decoder.BeamSearchDecoder(beam_width=100, top_paths=1)
# CTCPipeline
pipeline = asr.pipeline.ctc_pipeline.CTCPipeline(
alphabet=alphabet_en, features_extractor=features_extractor, model=model, optimizer=optimizer, decoder=decoder,
sample_rate=16000, mono=True, multi_gpu=multi_gpu
)
return pipeline
train_data = pd.read_csv('train_data.csv')
pipeline = get_config(feature_type = 'fbank', multi_gpu=False)
# train asr model
history = pipeline.fit(train_dataset=train_data, batch_size=128, epochs=500)
# history = pipeline.fit_generator(train_dataset = train_data, batch_size=32, epochs=500)
pipeline.save('./checkpoint')
Loaded pre-trained model has all components. The prediction can be invoked just by calling pipline.predict().
import pandas as pd
import deepasr as asr
import numpy as np
test_data = pd.read_csv('test_data.csv')
# get testing audio and transcript from dataset
index = np.random.randint(test_data.shape[0])
data = test_data.iloc[index]
test_file = data[0]
test_transcript = data[1]
# Test Audio file
print("Audio File:",test_file)
# Test Transcript
print("Audio Transcript:", test_transcript)
print("Transcript length:",len(test_transcript))
pipeline = asr.pipeline.load('./checkpoint')
print("Prediction", pipeline.predict(test_file))
import deepasr as asr
import numpy as np
test_data = pd.read_csv('test_data.csv')
# get testing audio and transcript from dataset
index = np.random.randint(test_data.shape[0])
data = test_data.iloc[index]
test_file = data[0]
test_transcript = data[1]
# Test Audio file
print("Audio File:",test_file)
# Test Transcript
print("Audio Transcript:", test_transcript)
print("Transcript length:",len(test_transcript))
pipeline = asr.pipeline.load('./checkpoint')
print("Prediction", pipeline.predict(test_file))
References
The fundamental repositories:
- Baidu - DeepSpeech2 - A PaddlePaddle implementation of DeepSpeech2 architecture for ASR
- NVIDIA - Toolkit for efficient experimentation with Speech Recognition, Text2Speech and NLP
- TensorFlow - The implementation of DeepSpeech2 model
- Mozilla - DeepSpeech - A TensorFlow implementation of Baidu's DeepSpeech architecture
- Espnet - End-to-End Speech Processing Toolkit
- Automatic Speech Recognition - Distill the Automatic Speech Recognition research
- Python Speech Features - Speech features for ASR including MFCCs and filterbank energies