bp500-base100k_voxpopuli: Wav2vec 2.0 with Brazilian Portuguese (BP) Dataset

This is a the demonstration of a fine-tuned Wav2vec model for Brazilian Portuguese using the following datasets:

• CETUC: contains approximately 145 hours of Brazilian Portuguese speech distributed among 50 male and 50 female speakers, each pronouncing approximately 1,000 phonetically balanced sentences selected from the CETEN-Folha corpus.
• Common Voice 7.0: is a project proposed by Mozilla Foundation with the goal to create a wide open dataset in different languages. In this project, volunteers donate and validate speech using the oficial site.
• Lapsbm: "Falabrasil - UFPA" is a dataset used by the Fala Brasil group to benchmark ASR systems in Brazilian Portuguese. Contains 35 speakers (10 females), each one pronouncing 20 unique sentences, totalling 700 utterances in Brazilian Portuguese. The audios were recorded in 22.05 kHz without environment control.
• Multilingual Librispeech (MLS): a massive dataset available in many languages. The MLS is based on audiobook recordings in public domain like LibriVox. The dataset contains a total of 6k hours of transcribed data in many languages. The set in Portuguese used in this work (mostly Brazilian variant) has approximately 284 hours of speech, obtained from 55 audiobooks read by 62 speakers.
• Multilingual TEDx: a collection of audio recordings from TEDx talks in 8 source languages. The Portuguese set (mostly Brazilian Portuguese variant) contains 164 hours of transcribed speech.
• Sidney (SID): contains 5,777 utterances recorded by 72 speakers (20 women) from 17 to 59 years old with fields such as place of birth, age, gender, education, and occupation;
• VoxForge: is a project with the goal to build open datasets for acoustic models. The corpus contains approximately 100 speakers and 4,130 utterances of Brazilian Portuguese, with sample rates varying from 16kHz to 44.1kHz.

These datasets were combined to build a larger Brazilian Portuguese dataset. All data was used for training except Common Voice dev/test sets, that were used for validation/test respectively. We also made test sets for all the gathered datasets.

Dataset	Train	Valid	Test
CETUC	94.0h	--	5.4h
Common Voice	37.8h	8.9h	9.5h
LaPS BM	0.8h	--	0.1h
MLS	161.0h	--	3.7h
Multilingual TEDx (Portuguese)	148.9h	--	1.8h
SID	7.2h	--	1.0h
VoxForge	3.9h	--	0.1h
Total	453.6h	8.9h	21.6h

The original model was fine-tuned using fairseq. This notebook uses a converted version of the original one. The link to the original fairseq model is available here.

Summary

	CETUC	CV	LaPS	MLS	SID	TEDx	VF	AVG
bp_500-base100k_voxpopuli (demonstration below)	0.142	0.201	0.052	0.224	0.102	0.317	0.048	0.155
bp_500-base100k_voxpopuli + 4-gram (demonstration below)	0.099	0.149	0.047	0.192	0.115	0.371	0.127	0.157

Transcription examples

Text	Transcription
qual o instagram dele	qualo está gramedele
o capitão foi expulso do exército porque era doido	o capitãl foi exposo do exército porque era doido
também por que não	também porque não
não existe tempo como o presente	não existe tempo como o presente
eu pulei para salvar rachel	eu pulei para salvar haquel
augusto cezar passos marinho	augusto cesa passoesmarinho

Demonstration

MODEL_NAME = "lgris/bp500-base100k_voxpopuli" 

Imports and dependencies

%%capture
!pip install torch==1.8.2+cu111 torchvision==0.9.2+cu111 torchaudio===0.8.2 -f https://download.pytorch.org/whl/lts/1.8/torch_lts.html
!pip install datasets
!pip install jiwer
!pip install transformers
!pip install soundfile
!pip install pyctcdecode
!pip install https://github.com/kpu/kenlm/archive/master.zip

import jiwer
import torchaudio
from datasets import load_dataset, load_metric
from transformers import (
    Wav2Vec2ForCTC,
    Wav2Vec2Processor,
)
from pyctcdecode import build_ctcdecoder
import torch
import re
import sys

Helpers

chars_to_ignore_regex = '[\,\?\.\!\;\:\"]'  # noqa: W605

def map_to_array(batch):
    speech, _ = torchaudio.load(batch["path"])
    batch["speech"] = speech.squeeze(0).numpy() 
    batch["sampling_rate"] = 16_000 
    batch["sentence"] = re.sub(chars_to_ignore_regex, '', batch["sentence"]).lower().replace("’", "'")
    batch["target"] = batch["sentence"]
    return batch

def calc_metrics(truths, hypos):
    wers = []
    mers = []
    wils = []
    for t, h in zip(truths, hypos):
        try:
            wers.append(jiwer.wer(t, h))
            mers.append(jiwer.mer(t, h))
            wils.append(jiwer.wil(t, h))
        except: # Empty string?
            pass
    wer = sum(wers)/len(wers)
    mer = sum(mers)/len(mers)
    wil = sum(wils)/len(wils)
    return wer, mer, wil

def load_data(dataset):
    data_files = {'test': f'{dataset}/test.csv'}
    dataset = load_dataset('csv', data_files=data_files)["test"]
    return dataset.map(map_to_array)

Model

class STT:

    def __init__(self, 
                 model_name, 
                 device='cuda' if torch.cuda.is_available() else 'cpu', 
                 lm=None):
        self.model_name = model_name
        self.model = Wav2Vec2ForCTC.from_pretrained(model_name).to(device)
        self.processor = Wav2Vec2Processor.from_pretrained(model_name)
        self.vocab_dict = self.processor.tokenizer.get_vocab()
        self.sorted_dict = {
            k.lower(): v for k, v in sorted(self.vocab_dict.items(), 
                                            key=lambda item: item[1])
        }
        self.device = device
        self.lm = lm
        if self.lm:            
            self.lm_decoder = build_ctcdecoder(
                list(self.sorted_dict.keys()),
                self.lm
            )

    def batch_predict(self, batch):
        features = self.processor(batch["speech"], 
                                  sampling_rate=batch["sampling_rate"][0], 
                                  padding=True, 
                                  return_tensors="pt")
        input_values = features.input_values.to(self.device)
        with torch.no_grad():
            logits = self.model(input_values).logits
        if self.lm:
            logits = logits.cpu().numpy()
            batch["predicted"] = []
            for sample_logits in logits:
                batch["predicted"].append(self.lm_decoder.decode(sample_logits))
        else:
            pred_ids = torch.argmax(logits, dim=-1)
            batch["predicted"] = self.processor.batch_decode(pred_ids)
        return batch

Download datasets

%%capture
!gdown --id 1HFECzIizf-bmkQRLiQD0QVqcGtOG5upI
!mkdir bp_dataset
!unzip bp_dataset -d bp_dataset/

%cd bp_dataset 

/content/bp_dataset

Tests

stt = STT(MODEL_NAME)

CETUC

ds = load_data('cetuc_dataset')
result = ds.map(stt.batch_predict, batched=True, batch_size=8) 
wer, mer, wil = calc_metrics(result["sentence"], result["predicted"])
print("CETUC WER:", wer)

CETUC WER: 0.1419179499917191

Common Voice

ds = load_data('commonvoice_dataset')
result = ds.map(stt.batch_predict, batched=True, batch_size=8) 
wer, mer, wil = calc_metrics(result["sentence"], result["predicted"])
print("CV WER:", wer)

CV WER: 0.20079950312040154

LaPS

ds = load_data('lapsbm_dataset')
result = ds.map(stt.batch_predict, batched=True, batch_size=8) 
wer, mer, wil = calc_metrics(result["sentence"], result["predicted"])
print("Laps WER:", wer)

Laps WER: 0.052780934343434324

MLS

ds = load_data('mls_dataset')
result = ds.map(stt.batch_predict, batched=True, batch_size=8) 
wer, mer, wil = calc_metrics(result["sentence"], result["predicted"])
print("MLS WER:", wer)

MLS WER: 0.22413887199364113

SID

ds = load_data('sid_dataset')
result = ds.map(stt.batch_predict, batched=True, batch_size=8) 
wer, mer, wil = calc_metrics(result["sentence"], result["predicted"])
print("Sid WER:", wer)

Sid WER: 0.1019041538671034

TEDx

ds = load_data('tedx_dataset')
result = ds.map(stt.batch_predict, batched=True, batch_size=8) 
wer, mer, wil = calc_metrics(result["sentence"], result["predicted"])
print("TEDx WER:", wer)

TEDx WER: 0.31711268778273327

VoxForge

ds = load_data('voxforge_dataset')
result = ds.map(stt.batch_predict, batched=True, batch_size=8) 
wer, mer, wil = calc_metrics(result["sentence"], result["predicted"])
print("VoxForge WER:", wer)

VoxForge WER: 0.04826433982683982

Tests with LM

!rm -rf ~/.cache
!gdown --id 1GJIKseP5ZkTbllQVgOL98R4yYAcIySFP  # trained with wikipedia
stt = STT(MODEL_NAME, lm='pt-BR-wiki.word.4-gram.arpa')
# !gdown --id 1dLFldy7eguPtyJj5OAlI4Emnx0BpFywg  # trained with bp
# stt = STT(MODEL_NAME, lm='pt-BR.word.4-gram.arpa')

Cetuc

ds = load_data('cetuc_dataset')
result = ds.map(stt.batch_predict, batched=True, batch_size=8) 
wer, mer, wil = calc_metrics(result["sentence"], result["predicted"])
print("CETUC WER:", wer)

CETUC WER: 0.099518615112877

Common Voice

ds = load_data('commonvoice_dataset')
result = ds.map(stt.batch_predict, batched=True, batch_size=8) 
wer, mer, wil = calc_metrics(result["sentence"], result["predicted"])
print("CV WER:", wer)

CV WER: 0.1488912889506362

LaPS

ds = load_data('lapsbm_dataset')
result = ds.map(stt.batch_predict, batched=True, batch_size=8) 
wer, mer, wil = calc_metrics(result["sentence"], result["predicted"])
print("Laps WER:", wer)

Laps WER: 0.047080176767676764

MLS

ds = load_data('mls_dataset')
result = ds.map(stt.batch_predict, batched=True, batch_size=8) 
wer, mer, wil = calc_metrics(result["sentence"], result["predicted"])
print("MLS WER:", wer)

MLS WER: 0.19220291966887196

SID

ds = load_data('sid_dataset')
result = ds.map(stt.batch_predict, batched=True, batch_size=8) 
wer, mer, wil = calc_metrics(result["sentence"], result["predicted"])
print("Sid WER:", wer)

Sid WER: 0.11535498771650306

TEDx

ds = load_data('tedx_dataset')
result = ds.map(stt.batch_predict, batched=True, batch_size=8) 
wer, mer, wil = calc_metrics(result["sentence"], result["predicted"])
print("TEDx WER:", wer)

TEDx WER: 0.3707890073539895

VoxForge

ds = load_data('voxforge_dataset')
result = ds.map(stt.batch_predict, batched=True, batch_size=8) 
wer, mer, wil = calc_metrics(result["sentence"], result["predicted"])
print("VoxForge WER:", wer)

VoxForge WER: 0.12682088744588746