Russo Discurso Aberto para Texto

Funções auxiliares / dependências

Construindo libsndfile

Uma maneira eficiente de ler arquivos opus em Python que não incorre em sobrecarga significativa é usar pysoundfile (um wrapper CFFI Python em torno de libsoundfile).

O suporte ao Opus foi implementado a montante, mas não foi lançado corretamente. Portanto, optamos por build personalizado + monkey patching.

Normalmente, você precisa executar isso em seu shell com acesso sudo:

apt-get update
apt-get install cmake autoconf autogen automake build-essential libasound2-dev \
libflac-dev libogg-dev libtool libvorbis-dev libopus-dev pkg-config -y

cd /usr/local/lib
git clone https://github.com/erikd/libsndfile.git
cd libsndfile
git reset --hard 49b7d61
mkdir -p build && cd build

cmake .. -DBUILD_SHARED_LIBS=ON
make && make install
cmake --build .

Funções auxiliares / dependências

Instale as bibliotecas a seguir:

pandas
numpy
scipy
tqdm
soundfile
librosa

Manifestos são arquivos csv com as seguintes colunas:

• Caminho para o áudio
• Caminho para o arquivo de texto
• Duração

Eles provaram ser o formato mais simples de acesso aos dados.

Para facilitar o uso, todos os manifestos já estão reenraizados. Todos os caminhos neles são relativos, você precisa fornecer uma pasta raiz.

# manifest utils
import os
import numpy as np
import pandas as pd
from tqdm import tqdm
from urllib.request import urlopen


def reroot_manifest(manifest_df,
                    source_path,
                    target_path):
    if source_path != '':
        manifest_df.wav_path = manifest_df.wav_path.apply(lambda x: x.replace(source_path,
                                                                              target_path))
        manifest_df.text_path = manifest_df.text_path.apply(lambda x: x.replace(source_path,
                                                                                target_path))
    else:
        manifest_df.wav_path = manifest_df.wav_path.apply(lambda x: os.path.join(target_path, x))
        manifest_df.text_path = manifest_df.text_path.apply(lambda x: os.path.join(target_path, x))    
    return manifest_df


def save_manifest(manifest_df,
                  path,
                  domain=False):
    if domain:
        assert list(manifest_df.columns) == ['wav_path', 'text_path', 'duration', 'domain']
    else:
        assert list(manifest_df.columns) == ['wav_path', 'text_path', 'duration']

    manifest_df.reset_index(drop=True).sort_values(by='duration',
                                                   ascending=True).to_csv(path,
                                                                          sep=',',
                                                                          header=False,
                                                                          index=False)
    return True


def read_manifest(manifest_path,
                  domain=False):
    if domain:
        return pd.read_csv(manifest_path,
                        names=['wav_path',
                               'text_path',
                               'duration',
                               'domain'])
    else:
        return pd.read_csv(manifest_path,
                        names=['wav_path',
                               'text_path',
                               'duration'])


def check_files(manifest_df,
                domain=False):
    orig_len = len(manifest_df)
    if domain:
        assert list(manifest_df.columns) == ['wav_path', 'text_path', 'duration']
    else:
        assert list(manifest_df.columns) == ['wav_path', 'text_path', 'duration', 'domain']
    wav_paths = list(manifest_df.wav_path.values)
    text_path = list(manifest_df.text_path.values)

    omitted_wavs = []
    omitted_txts = []

    for wav_path, text_path in zip(wav_paths, text_path):
        if not os.path.exists(wav_path):
            print('Dropping {}'.format(wav_path))
            omitted_wavs.append(wav_path)
        if not os.path.exists(text_path):
            print('Dropping {}'.format(text_path))
            omitted_txts.append(text_path)

    manifest_df = manifest_df[~manifest_df.wav_path.isin(omitted_wavs)]
    manifest_df = manifest_df[~manifest_df.text_path.isin(omitted_txts)]
    final_len = len(manifest_df)

    if final_len != orig_len:
        print('Removed {} lines'.format(orig_len-final_len))
    return manifest_df


def plain_merge_manifests(manifest_paths,
                          MIN_DURATION=0.1,
                          MAX_DURATION=100):

    manifest_df = pd.concat([read_manifest(_)
                             for _ in manifest_paths])
    manifest_df = check_files(manifest_df)

    manifest_df_fit = manifest_df[(manifest_df.duration>=MIN_DURATION) &
                                  (manifest_df.duration<=MAX_DURATION)]

    manifest_df_non_fit = manifest_df[(manifest_df.duration<MIN_DURATION) |
                                      (manifest_df.duration>MAX_DURATION)]

    print(f'Good hours: {manifest_df_fit.duration.sum() / 3600:.2f}')
    print(f'Bad hours: {manifest_df_non_fit.duration.sum() / 3600:.2f}')

    return manifest_df_fit


def save_txt_file(wav_path, text):
    txt_path = wav_path.replace('.wav','.txt')
    with open(txt_path, "w") as text_file:
        print(text, file=text_file)
    return txt_path


def read_txt_file(text_path):
    #with open(text_path, 'r') as file:
    response = urlopen(text_path)
    file = response.readlines()
    for i in range(len(file)):
        file[i] = file[i].decode('utf8')
    return file 

def create_manifest_from_df(df, domain=False):
    if domain:
        columns = ['wav_path', 'text_path', 'duration', 'domain']
    else:
        columns = ['wav_path', 'text_path', 'duration']
    manifest = df[columns]
    return manifest


def create_txt_files(manifest_df):
    assert 'text' in manifest_df.columns
    assert 'wav_path' in manifest_df.columns
    wav_paths, texts = list(manifest_df['wav_path'].values), list(manifest_df['text'].values)
    # not using multiprocessing for simplicity
    txt_paths = [save_txt_file(*_) for _ in tqdm(zip(wav_paths, texts), total=len(wav_paths))]
    manifest_df['text_path'] = txt_paths
    return manifest_df


def replace_encoded(text):
    text = text.lower()
    if '2' in text:
        text = list(text)
        _text = []
        for i,char in enumerate(text):
            if char=='2':
                try:
                    _text.extend([_text[-1]])
                except:
                    print(''.join(text))
            else:
                _text.extend([char])
        text = ''.join(_text)
    return text

# reading opus files
import os
import soundfile as sf



# Fx for soundfile read/write functions
def fx_seek(self, frames, whence=os.SEEK_SET):
    self._check_if_closed()
    position = sf._snd.sf_seek(self._file, frames, whence)
    return position


def fx_get_format_from_filename(file, mode):
    format = ''
    file = getattr(file, 'name', file)
    try:
        format = os.path.splitext(file)[-1][1:]
        format = format.decode('utf-8', 'replace')
    except Exception:
        pass
    if format == 'opus':
        return 'OGG'
    if format.upper() not in sf._formats and 'r' not in mode:
        raise TypeError("No format specified and unable to get format from "
                        "file extension: {0!r}".format(file))
    return format


#sf._snd = sf._ffi.dlopen('/usr/local/lib/libsndfile/build/libsndfile.so.1.0.29')
sf._subtypes['OPUS'] = 0x0064
sf.SoundFile.seek = fx_seek
sf._get_format_from_filename = fx_get_format_from_filename


def read(file, **kwargs):
    return sf.read(file, **kwargs)


def write(file, data, samplerate, **kwargs):
    return sf.write(file, data, samplerate, **kwargs)

# display utils
import gc
from IPython.display import HTML, Audio, display_html
pd.set_option('display.max_colwidth', 3000)
#Prepend_path is set to read directly from Azure. To read from local replace below string with path to the downloaded dataset files
prepend_path = 'https://azureopendatastorage.blob.core.windows.net/openstt/ru_open_stt_opus_unpacked/'


def audio_player(audio_path):
    return '<audio preload="none" controls="controls"><source src="{}" type="audio/wav"></audio>'.format(audio_path)

def display_manifest(manifest_df):
    display_df = manifest_df
    display_df['wav'] = [audio_player(prepend_path+path) for path in display_df.wav_path]
    display_df['txt'] = [read_txt_file(prepend_path+path) for path in tqdm(display_df.text_path)]
    audio_style = '<style>audio {height:44px;border:0;padding:0 20px 0px;margin:-10px -20px -20px;}</style>'
    display_df = display_df[['wav','txt', 'duration']]
    display(HTML(audio_style + display_df.to_html(escape=False)))
    del display_df
    gc.collect()

Jogar com um conjunto de dados

Reproduzir uma amostra de ficheiros

A maioria dos navegadores de plataformas suporta reprodução de áudio nativa. Assim, podemos usar players de áudio HTML5 para visualizar nossos dados.

manifest_df = read_manifest(prepend_path +'/manifests/public_series_1.csv')
#manifest_df = reroot_manifest(manifest_df,
                              #source_path='',
                              #target_path='../../../../../nvme/stt/data/ru_open_stt/')

sample = manifest_df.sample(n=20)
display_manifest(sample)

Ler um ficheiro

!ls ru_open_stt_opus/manifests/*.csv

Alguns exemplos mostrando como ler melhor arquivos wav e opus.

Scipy é o mais rápido para wav. Pysoundfile é o melhor geral para opus.

%matplotlib inline

import librosa
from scipy.io import wavfile
from librosa import display as ldisplay
from matplotlib import pyplot as plt

Ler um wav

manifest_df = read_manifest(prepend_path +'manifests/asr_calls_2_val.csv')
#manifest_df = reroot_manifest(manifest_df,
                              #source_path='',
                              #target_path='../../../../../nvme/stt/data/ru_open_stt/')

sample = manifest_df.sample(n=5)
display_manifest(sample)

from io import BytesIO

wav_path = sample.iloc[0].wav_path
response = urlopen(prepend_path+wav_path)
data = response.read()
sr, wav = wavfile.read(BytesIO(data))
wav.astype('float32')
absmax = np.max(np.abs(wav))
wav =  wav / absmax

# shortest way to plot a spectrogram
D = librosa.amplitude_to_db(np.abs(librosa.stft(wav)), ref=np.max)
plt.figure(figsize=(12, 6))
ldisplay.specshow(D, y_axis='log')
plt.colorbar(format='%+2.0f dB')
plt.title('Log-frequency power spectrogram')
# shortest way to plot an envelope
plt.figure(figsize=(12, 6))
ldisplay.waveplot(wav, sr=sr, max_points=50000.0, x_axis='time', offset=0.0, max_sr=1000, ax=None)

Ler opus

manifest_df = read_manifest(prepend_path +'manifests/asr_public_phone_calls_2.csv')
#manifest_df = reroot_manifest(manifest_df,
                              #source_path='',
                              #target_path='../../../../../nvme/stt/data/ru_open_stt/')

sample = manifest_df.sample(n=5)
display_manifest(sample)

opus_path = sample.iloc[0].wav_path
response = urlopen(prepend_path+opus_path)
data = response.read()
wav, sr = sf.read(BytesIO(data))
wav.astype('float32')
absmax = np.max(np.abs(wav))
wav =  wav / absmax

# shortest way to plot a spectrogram
D = librosa.amplitude_to_db(np.abs(librosa.stft(wav)), ref=np.max)
plt.figure(figsize=(12, 6))
ldisplay.specshow(D, y_axis='log')
plt.colorbar(format='%+2.0f dB')
plt.title('Log-frequency power spectrogram')
# shortest way to plot an envelope
plt.figure(figsize=(12, 6))
ldisplay.waveplot(wav, sr=sr, max_points=50000.0, x_axis='time', offset=0.0, max_sr=1000, ax=None)