class ConformerTamilASR(object):
"""
Conformer S based ASR model
"""
def __init__(self, path='ConformerS.h5'):
# fetch and load the config of the model
config = Config('tamil_tech/configs/conformer_new_config.yml', learning=True)
# load speech and text featurizers
speech_featurizer = TFSpeechFeaturizer(config.speech_config)
text_featurizer = CharFeaturizer(config.decoder_config)
# check if model already exists in given path, else download the model in the given path
if os.path.exists(path):
pass
else:
print("Downloading Model...")
file_id = config.file_id
download_file_from_google_drive(file_id, path)
print("Downloaded Model Successfully...")
# load model using config
self.model = Conformer(**config.model_config, vocabulary_size=text_featurizer.num_classes)
# set shape of the featurizer and build the model
self.model._build(speech_featurizer.shape)
# load weights of the model
self.model.load_weights(path, by_name=True)
# display model summary
self.model.summary(line_length=120)
# set featurizers for the model
self.model.add_featurizers(speech_featurizer, text_featurizer)
print("Loaded Model...!")
def read_raw_audio(self, audio, sample_rate=16000):
# if audio path is given, load audio using librosa
if isinstance(audio, str):
wave, _ = librosa.load(os.path.expanduser(audio), sr=sample_rate)
# if audio file is in bytes, use soundfile to read audio
elif isinstance(audio, bytes):
wave, sr = sf.read(io.BytesIO(audio))
# if audio is stereo, convert it to mono
try:
if wave.shape[1] >= 2:
wave = np.transpose(wave)[0][:]
except:
pass
# get loaded audio as numpy array
wave = np.asfortranarray(wave)
# resampel to 16000 kHz
if sr != sample_rate:
wave = librosa.resample(wave, sr, sample_rate)
# if numpy array, return audio
elif isinstance(audio, np.ndarray):
return audio
else:
raise ValueError("input audio must be either a path or bytes")
return wave
def bytes_to_string(self, array: np.ndarray, encoding: str = "utf-8"):
# decode text array with utf-8 encoding
return [transcript.decode(encoding) for transcript in array]
def infer(self, path, greedy=True, return_text=False):
# read the audio
signal = self.read_raw_audio(path)
# expand dims to process for a single prediction
signal = tf.expand_dims(self.model.speech_featurizer.tf_extract(signal), axis=0)
# predict greedy
if greedy:
pred = self.model.recognize(features=signal)
else:
# preidct using beam search and language model
pred = self.model.recognize_beam(features=signal, lm=True)
if return_text:
# return predicted transcription
return self.bytes_to_string(pred.numpy())[0]
# return predicted transcription
print(self.bytes_to_string(pred.numpy())[0], end=' ')
import numpy as np
np.random.seed(0)
tf.random.set_seed(0)
if args.filename.endswith('.wav'):
signal = read_raw_audio(args.filename)
# features = speech_featurizer.tf_extract(signal)
features = speech_featurizer.extract(signal)
features = tf.constant(features)
else:
features = np.load(args.filename).reshape([-1, 80, 1])
features = tf.constant(features)
input_length = get_reduced_length(tf.shape(features)[0], conformer.time_reduction_factor)
if args.beam_width:
transcript = conformer.recognize_beam(features[None, ...], input_length[None, ...])
print("Transcript:", transcript[0].numpy().decode("UTF-8"))
elif args.timestamp:
transcript, stime, etime, _, _ = conformer.recognize_tflite_with_timestamp(
signal, tf.constant(text_featurizer.blank, dtype=tf.int32), conformer.predict_net.get_initial_state())
print("Transcript:", transcript)
print("Start time:", stime)
print("End time:", etime)
else:
if args.filename.endswith('.wav'):
transcript, _, _ = conformer.recognize_tflite(
signal, tf.constant(text_featurizer.blank, dtype=tf.int32), conformer.predict_net.get_initial_state())
print("Transcript:", tf.strings.unicode_encode(transcript, "UTF-8").numpy().decode("UTF-8"))
else:
encoded = conformer.encoder_inference(features)
hypothesis = conformer._perform_greedy(
from tensorflow_asr.configs.config import Config
from tensorflow_asr.featurizers.speech_featurizers import read_raw_audio
from tensorflow_asr.featurizers.speech_featurizers import TFSpeechFeaturizer
from tensorflow_asr.featurizers.text_featurizers import CharFeaturizer, SubwordFeaturizer
from tensorflow_asr.models.conformer import Conformer
config = Config(args.config, learning=False)
speech_featurizer = TFSpeechFeaturizer(config.speech_config)
if args.subwords and os.path.exists(args.subwords):
print("Loading subwords ...")
text_featurizer = SubwordFeaturizer.load_from_file(config.decoder_config, args.subwords)
else:
text_featurizer = CharFeaturizer(config.decoder_config)
text_featurizer.decoder_config.beam_width = args.beam_width
# build model
conformer = Conformer(**config.model_config, vocabulary_size=text_featurizer.num_classes)
conformer._build(speech_featurizer.shape)
conformer.load_weights(args.saved, by_name=True)
conformer.summary(line_length=120)
conformer.add_featurizers(speech_featurizer, text_featurizer)
signal = read_raw_audio(args.filename)
if (args.beam_width):
transcript = conformer.recognize_beam(signal[None, ...])
else:
transcript = conformer.recognize(signal[None, ...])
tf.print("Transcript:", transcript[0])
