Speech-to-Text HuggingFace + pyctcdecode#

Finetuned hyperlocal languages on pretrained HuggingFace models + pyctcdecode with KenLM, https://huggingface.co/mesolitica

This tutorial is available as an IPython notebook at malaya-speech/example/stt-huggingface-pyctcdecode.

This module is not language independent, so it not save to use on different languages. Pretrained models trained on hyperlocal languages.

This is an application of malaya-speech Pipeline, read more about malaya-speech Pipeline at malaya-speech/example/pipeline.

Required Tensorflow >= 2.0 due to group convolution is not available for Tensorflow 1.

[1]:

import malaya_speech
import numpy as np
from malaya_speech import Pipeline

Install pyctcdecode#

From PYPI#

pip3 install pyctcdecode==0.1.0 pypi-kenlm==0.1.20210121

From source#

Check https://github.com/kensho-technologies/pyctcdecode how to build from source incase there is no available wheel for your operating system.

Building from source should only take a few minutes.

Benefit#

  1. pyctcdecode accurate than ctc-decoders for certain cases, but slower than pyctcdecode.

  2. pip install and done, no need to compile.

List available HuggingFace model#

[2]:

malaya_speech.stt.available_huggingface()

[2]:
CER CER-LM Language Size (MB) WER WER-LM
mesolitica/wav2vec2-xls-r-300m-mixed 0.048105 0.041196 [malay, singlish, mandarin] 1180 0.13222 0.098802

Load HuggingFace model#

def huggingface(model: str = 'mesolitica/wav2vec2-xls-r-300m-mixed', **kwargs):
    """
    Load Finetuned models from HuggingFace. Required Tensorflow >= 2.0.

    Parameters
    ----------
    model : str, optional (default='mesolitica/wav2vec2-xls-r-300m-mixed')
        Model architecture supported. Allowed values:

        * ``'mesolitica/wav2vec2-xls-r-300m-mixed'`` - wav2vec2 XLS-R 300M finetuned on (Malay + Singlish + Mandarin) languages.

    Returns
    -------
    result : malaya_speech.model.huggingface.CTC class
    """

[3]:

model = malaya_speech.stt.huggingface(model = 'mesolitica/wav2vec2-xls-r-300m-mixed')

Load sample#

[4]:

ceramah, sr = malaya_speech.load('speech/khutbah/wadi-annuar.wav')
record1, sr = malaya_speech.load('speech/record/savewav_2020-11-26_22-36-06_294832.wav')
record2, sr = malaya_speech.load('speech/record/savewav_2020-11-26_22-40-56_929661.wav')
singlish0, sr = malaya_speech.load('speech/singlish/singlish0.wav')
singlish1, sr = malaya_speech.load('speech/singlish/singlish1.wav')
singlish2, sr = malaya_speech.load('speech/singlish/singlish2.wav')
mandarin0, sr = malaya_speech.load('speech/mandarin/597.wav')
mandarin1, sr = malaya_speech.load('speech/mandarin/584.wav')
mandarin2, sr = malaya_speech.load('speech/mandarin/509.wav')

Predict logits#

def predict_logits(self, inputs, norm_func=softmax):
    """
    Predict logits from inputs.

    Parameters
    ----------
    input: List[np.array]
        List[np.array] or List[malaya_speech.model.frame.Frame].
    norm_func: Callable, optional (default=malaya.utils.activation.softmax)


    Returns
    -------
    result: List[np.array]
    """

[5]:

%%time

logits = model.predict_logits([ceramah, record1, record2])

CPU times: user 35.2 s, sys: 16.7 s, total: 51.9 s
Wall time: 9.79 s

[6]:

logits.shape

[6]:

(3, 499, 40)

Load ctc-decoders#

I will use dump-combined for this example.

[7]:

from pyctcdecode import build_ctcdecoder
from malaya_speech.utils.char import HF_CTC_VOCAB
import kenlm

[8]:

lm = malaya_speech.language_model.kenlm(model = 'dump-combined')

[42]:

kenlm_model = kenlm.Model(lm)
decoder = build_ctcdecoder(
    HF_CTC_VOCAB + ['_'],
    kenlm_model,
    alpha=0.2,
    beta=1.0,
    ctc_token_idx=len(HF_CTC_VOCAB)
)

[34]:

len(HF_CTC_VOCAB)

[34]:

39

[43]:

out = decoder.decode_beams(logits[0], prune_history=True)
d_lm, lm_state, timesteps, logit_score, lm_score = out[0]
d_lm

[43]:

'jadi dalam perjalanan ini dunia yang susah ini ketika nabi mengajar muaz bin jabal tadi ni allah ma ini'