Spring over navigation

Russian Open Speech To Text

Speech to Text Russian Open STT

En samling taleeksempler, der er afledt fra forskellige lydkilder. Datasættet indeholder korte lydklip på russisk.

Det er muligvis det største russiske tale til tekst-datasæt til dato:

  • ~16 mio. taleudtryk;
  • ~20.000 timer;
  • 2,3 TB (ukomprimeret i .wav-format i int16), 356 G i .opus;
  • Nu er alle filer transformeret til opus med undtagelse af valideringsdatasæt;

Det primære formål med datasættet er at træne tale til tekst-modeller.

Datasætsammensætning

Datasætstørrelsen er givet for .wav filer.

Datasæt Taleudtryk Timer GB Sek./tegn Kommentar Anmærkning Kvalitet/støj
radio_v4 (*) 7.603.192 10.430 1.195 5s/68 Radio Juster 95 %/frisk
public_speech (*) 1.700.060 2.709 301 6s/79 Offentlig tale Juster 95 %/frisk
audiobook_2 1.149.404 1.511 162 5s/56 Bøger Juster 95 %/frisk
radio_2 651.645 1.439 154 8s/110 Radio Juster 95 %/frisk
public_youtube1120 1.410.979 1.104 237 3s/34 YouTube Undertekster 95%/~frisk
public_youtube700 759.483 701 75 3s/43 YouTube Undertekster 95%/~frisk
tts_russian_addresses 1.741.838 754 81 2s/20 Adresser TTS 4-stemmer 100 %/frisk
asr_public_phone_calls_2 603.797 601 66 4s/37 Telefonopkald ASR 70 %/støjende
public_youtube1120_hq 369.245 291 31 3s/37 YouTube HQ Undertekster 95%/~frisk
asr_public_phone_calls_1 233.868 211 23 3s/29 Telefonopkald ASR 70 %/støjende
radio_v4_add (*) 92.679 157 18 6s/80 Radio Juster 95 %/frisk
asr_public_stories_2 78.186 78 9 4s/43 Bøger ASR 80 %/frisk
asr_public_stories_1 46.142 38 4 3s/30 Bøger ASR 80 %/frisk
public_series_1 20.243 17 2 3s/38 YouTube Undertekster 95%/~frisk
asr_calls_2_val 12.950 7.7 2 2s/34 Telefonopkald Manual annotation 99 %/frisk
public_lecture_1 6.803 6 1 3s/47 Lektioner Undertekster 95 %/frisk
buriy_audiobooks_2_val 7.850 4.9 1 2s/31 Bøger Manual annotation 99 %/frisk
public_youtube700_val 7.311 4.5 1 2s/35 YouTube Manual annotation 99 %/frisk

(*) Der leveres kun en dataprøve med txt-filerne.

Annotationsmetodik

Datasættet er sammensat ved hjælp af åbne kilder. Lange sekvenser er opdelt i lyddele ved hjælp af registrering af stemmeaktivitet og justering. Nogle lydtyper er annoteret automatisk og bekræftet statistisk/ved hjælp af heuristik.

Datamængder og opdateringsfrekvens

Den samlede størrelse af hele datasættet er 350 GB. Den samlede størrelse af datasættet med offentligt delte mærkater er 130 GB.

Det er ikke sandsynligt, at selve datasættet opdateres for at være bagudkompatibelt. Følg det oprindelige lager i forbindelse med benchmarks og eksklusionsfiler.

Nye domæner og sprog kan tilføjes i fremtiden.

Lydnormalisering

Alle filer normaliseres for at sikre lettere/hurtigere kørselsforøgelse og -behandling som følger:

  • Konverteret til mono, hvis det er nødvendigt;
  • Konverteret til en samplingfrekvens på 16 kHz, hvis det er nødvendigt;
  • Gemt som 16-bit heltal;
  • Konverteret til OPUS;

På disk-databasemetodik

Hver enkelt lydfil (wav, binary) er hashkodet. Hashkoden bruges til at oprette et mappehierarki for at sikre en optimal fs-handling.

target_format = 'wav' wavb = wav.tobytes() f_hash = hashlib.sha1(wavb).hexdigest() store_path = Path(root_folder, f_hash[0], f_hash[1:3], f_hash[3:15] + '.' + target_format)
Downloads

Datasættet leveres i to former:

  • Arkiver, der er tilgængelige via Azure Blob Storage og/eller direkte links;
  • Originale filer er tilgængelige via Azure Blob Storage;

Alt gemmes i https://azureopendatastorage.blob.core.windows.net/openstt/

Mappestruktur:

└── ru_open_stt_opus <= archived folders │ │ │ ├── archives │ │ ├── asr_calls_2_val.tar.gz <= tar.gz archives with opus and wav files │ │ │ ... <= see the below table for enumeration │ │ └── tts_russian_addresses_rhvoice_4voices.tar.gz │ │ │ └── manifests │ ├── asr_calls_2_val.csv <= csv files with wav_path, text_path, duration (see notebooks) │ │ ... │ └── tts_russian_addresses_rhvoice_4voices.csv └── ru_open_stt_opus_unpacked <= a separate folder for each uploaded domain ├── public_youtube1120 │ ├── 0 <= see "On disk DB methodology" for details │ ├── 1 │ │ ├── 00 │ │ │ ... │ │ └── ff │ │ ├── *.opus <= actual files │ │ └── *.txt │ │ ... │ └── f ├── public_youtube1120_hq ├── public_youtube700_val ├── asr_calls_2_val ├── radio_2 ├── private_buriy_audiobooks_2 ├── asr_public_phone_calls_2 ├── asr_public_stories_2 ├── asr_public_stories_1 ├── public_lecture_1 ├── asr_public_phone_calls_1 ├── public_series_1 └── public_youtube700
Datasæt GB, wav GB, arkiv Arkiv Kilde Manifest
Oplær
Eksempel på radio og offentlig tale - 11,4 opus+txt - manifest
audiobook_2 162 25,8 opus+txt Internet + justering manifest
radio_2 154 24,6 opus+txt Radio manifest
public_youtube1120 237 19,0 opus+txt YouTube-videoer manifest
asr_public_phone_calls_2 66 9.4 opus+txt Internet + ASR manifest
public_youtube1120_hq 31 4,9 opus+txt YouTube-videoer manifest
asr_public_stories_2 9 1.4 opus+txt Internet + justering manifest
tts_russian_addresses_rhvoice_4voices 80,9 12,9 opus+txt TTS manifest
public_youtube700 75,0 12,2 opus+txt YouTube-videoer manifest
asr_public_phone_calls_1 22,7 3.2 opus+txt Internet + ASR manifest
asr_public_stories_1 4.1 0.7 opus+txt Offentlige historier manifest
public_series_1 1.9 0.3 opus+txt Offentlig serie manifest
public_lecture_1 0.7 0.1 opus+txt Internet + manual manifest
Val
asr_calls_2_val 2 0,8 wav+txt Internet manifest
buriy_audiobooks_2_val 1 0,5 wav+txt Bøger + manual manifest
public_youtube700_val 2 0.13 wav+txt YouTube-videoer + manual manifest
Download instruktioner

Direkte

Se her – https://github.com/snakers4/open_stt#download-instructions

Via opstilling af Azure Blob Storage

Se den notesbog, der findes på fanen “Dataadgang”

Kontakter

Hvis du har brug for hjælp eller har spørgsmål om dataene, kan du kontakte dataopretter(ne) på aveysov@gmail.com

Licens

Denne licens gør det muligt for genbrugere at distribuere, blande, implementere og bygge videre på materialet på et vilkårligt medie eller i et hvilket som helst format udelukkende til ikke-kommercielle formål og kun så længe, der henvises til opretteren. Det omfatter følgende elementer:
* BY – Der skal henvises til opretteren
* NC – Materialet må kun bruges til ikke-kommercielle formål

CC-BY-NC og kommerciel anvendelse er tilladt efter aftale med opretterne af datasættet.

Referencer/fortsat læsning

Originalt datasæt

  • https://github.com/snakers4/open_stt

Engelske artikler

  • https://thegradient.pub/towards-an-imagenet-moment-for-speech-to-text/
  • https://thegradient.pub/a-speech-to-text-practitioners-criticisms-of-industry-and-academia/

Kinesiske artikler

  • https://www.infoq.cn/article/4u58WcFCs0RdpoXev1E2

Russiske artikler

  • https://habr.com/ru/post/494006/
  • https://habr.com/ru/post/474462/

Access

Available inWhen to use
Azure Notebooks

Quickly explore the dataset with Jupyter notebooks hosted on Azure or your local machine.

Select your preferred service:

Azure Notebooks

Azure Notebooks

Package: Language: Python

Helper functions / dependencies

Building libsndfile

The best efficient way to read opus files in python (the we know of) that does incur any significant overhead is to use pysoundfile (a python CFFI wrapper around libsoundfile).

When this solution was being researched the community had been waiting for a major libsoundfile release for some time.

Opus support has been implemented some time ago upstream, but it has not been properly released. Therefore we opted for a custom build + monkey patching.

At the time when you read / use this - probably there will be decent / proper builds of libsndfile.

Please replace with your faviourite tool if there is one.

Typically, you need to run this in your shell with sudo access:

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 .

Helper functions / dependencies

Install the following libraries (versions do not matter much):

pandas
numpy
scipy
tqdm
soundfile
librosa

Depending on how this notebook is run, this sometimes can be as easy as (if, for example your miniconda is not installed under root):

In [ ]:
!pip install numpy
!pip install tqdm
!pip install scipy
!pip install pandas
!pip install soundfile
!pip install librosa
!pip install azure-storage-blob

Manifests are just csv files with the following columns:

  • Path to audio
  • Path to text file
  • Duration

They proved to be the most simple / helpful format of accessing data.

For ease of use all the manifests are already rerooted, i.e. all paths in them are relative and you just need to add a root folder.

In [1]:
# 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
In [2]:
# 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)
In [3]:
# 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()

Play with a dataset

Play a sample of files

On most platforms browsers usually support native audio playback.

So we can leverage HTML5 audio players to view our data.

In [4]:
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/')
In [5]:
sample = manifest_df.sample(n=20)
display_manifest(sample)
100%|██████████| 20/20 [00:07<00:00,  2.66it/s]
wav txt duration
5963 [пожалуйста прости всё в порядке\n] 2.48
19972 [хотелось бы хотя бы разок глазком на неё посмотреть раз такое дело\n] 5.68
15555 [они с егерем на след напали до инспектора не дозвониться\n] 3.84
430 [что то случилось\n] 1.36
4090 [так давай опаздываем\n] 2.16
18590 [да саид слушаю тебя троих нашли а в полётном листе\n] 4.60
17734 [надо сначала самому серьёзным человеком стать понимаешь\n] 4.32
978 [вот что случилось\n] 1.56
13269 [да паш юль пожалуйста не делай глупостей\n] 3.48
4957 [полусладкое или сухое\n] 2.32
1913 [ищи другую машину\n] 1.80
10522 [гражданин финн не зная что я полицейский\n] 3.08
9214 [ты чего трубку не берёшь я же переживаю\n] 2.88
10014 [я не окажу сопротивления я без оружия\n] 3.00
8351 [звони партнёру пусть он напишет\n] 2.80
3818 [ну что пойдём обсудим\n] 2.12
11097 [вы простите понимаете все об этом знают\n] 3.16
2989 [какие уж разводки\n] 2.00
12229 [я получается какой то диспетчер а не напарник\n] 3.28
5348 [я же тебе сказала никакой карелии\n] 2.40

Read a file

In [ ]:
!ls ru_open_stt_opus/manifests/*.csv

A couple of simplistic examples showing how to best read wav and opus files.

Scipy is the fastest for wav, pysoundfile is the best overall for opus.

In [6]:
%matplotlib inline

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

Read a wav

In [7]:
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/')
In [8]:
sample = manifest_df.sample(n=5)
display_manifest(sample)
100%|██████████| 5/5 [00:01<00:00,  2.61it/s]
wav txt duration
7802 [это же позитивные новости не негативные\n] 2.01
3590 [белый цветочек\n] 1.17
10594 [какое отношение имеет ваша пенсия к моему отделению\n] 3.14
4630 [есть есть видео\n] 1.35
468 [что ещё раз\n] 0.62
In [9]:
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
In [10]:
# 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)
Out[10]:
<matplotlib.collections.PolyCollection at 0x7fdf62f7e8d0>

Read opus

In [11]:
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/')
In [12]:
sample = manifest_df.sample(n=5)
display_manifest(sample)
100%|██████████| 5/5 [00:02<00:00,  2.24it/s]
wav txt duration
5018 [а вы кто\n] 0.96
143473 [пьеса дружбы нету\n] 1.86
272155 [не знаю где находится\n] 2.64
334225 [ты куда звонишь то куда ты звонишь ты знаешь\n] 3.12
143789 [помощник дежурного\n] 1.86
In [13]:
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
In [14]:
# 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)
Out[14]:
<matplotlib.collections.PolyCollection at 0x7fdf62f8ee10>
In [ ]: