• Silero Models
  • Installation and Basics
  • Speech-To-Text
    • Dependencies
    • PyTorch
    • ONNX
    • TensorFlow
  • Text-To-Speech
    • Models and Speakers
    • Dependencies
    • PyTorch
    • Standalone Use
    • SSML
    • Cyrillic languages
    • Indic languages
  • Text-Enhancement
    • Dependencies
    • Standalone Use
  • Denoise
    • Models
    • Dependencies
    • PyTorch
    • Standalone Use
  • FAQ
    • Wiki
    • Performance and Quality
    • Adding new Languages
  • Contact
    • Get in Touch
    • Commercial Inquiries
  • Citations
  • Further reading
    • English
    • Chinese
    • Russian
  • Donations

Silero Models

Silero Models: pre-trained enterprise-grade STT / TTS models and benchmarks.

Enterprise-grade STT made refreshingly simple (seriously, see benchmarks). We provide quality comparable to Google's STT (and sometimes even better) and we are not Google.

As a bonus:

• No Kaldi;
• No compilation;
• No 20-step instructions;

Also we have published TTS models that satisfy the following criteria:

• One-line usage;
• A large library of voices;
• A fully end-to-end pipeline;
• Natural-sounding speech;
• No GPU or training required;
• Minimalism and lack of dependencies;
• Faster than real-time on one CPU thread (!!!);
• Support for 16kHz and 8kHz out of the box;

Also we have published a model for text repunctuation and recapitalization that:

• Inserts capital letters and basic punctuation marks, e.g., dots, commas, hyphens, question marks, exclamation points, and dashes (for Russian);
• Works for 4 languages (Russian, English, German, and Spanish) and can be extended;
• Domain-agnostic by design and not based on any hard-coded rules;
• Has non-trivial metrics and succeeds in the task of improving text readability;

Installation and Basics

You can basically use our models in 3 flavours:

• Via PyTorch Hub: torch.hub.load();
• Via pip: pip install silero and then import silero;
• Via caching the required models and utils manually and modifying if necessary;

Models are downloaded on demand both by pip and PyTorch Hub. If you need caching, do it manually or via invoking a necessary model once (it will be downloaded to a cache folder). Please see these docs for more information.

PyTorch Hub and pip package are based on the same code. All of the torch.hub.load examples can be used with the pip package via this basic change:

# before
torch.hub.load(repo_or_dir='snakers4/silero-models',
               model='silero_stt',  # or silero_tts or silero_te
               **kwargs)

# after
from silero import silero_stt, silero_tts, silero_te
silero_stt(**kwargs)

Speech-To-Text

All of the provided models are listed in the models.yml file. Any metadata and newer versions will be added there.

Currently we provide the following checkpoints:

	PyTorch	ONNX	Quantization	Quality	Colab
English (en_v6)	✔️	✔️	✔️	link
English (en_v5)	✔️	✔️	✔️	link
German (de_v4)	✔️	✔️	⌛	link
English (en_v3)	✔️	✔️	✔️	link
German (de_v3)	✔️	⌛	⌛	link
German (de_v1)	✔️	✔️	⌛	link
Spanish (es_v1)	✔️	✔️	⌛	link
Ukrainian (ua_v3)	✔️	✔️	✔️	N/A

Model flavours:

	jit	jit	jit	jit	jit_q	jit_q	onnx	onnx	onnx	onnx
	xsmall	small	large	xlarge	xsmall	small	xsmall	small	large	xlarge
English en_v6		✔️		✔️		✔️		✔️		✔️
English en_v5		✔️		✔️		✔️		✔️		✔️
English en_v4_0			✔️						✔️
English en_v3	✔️	✔️	✔️		✔️	✔️	✔️	✔️	✔️
German de_v4			✔️						✔️
German de_v3			✔️
German de_v1		✔️					✔️
Spanish es_v1		✔️					✔️
Ukrainian ua_v3		✔️			✔️		✔️

Dependencies

• All examples:
  • torch, 1.8+ (used to clone the repo in TensorFlow and ONNX examples), breaking changes for versions older than 1.6
  • torchaudio, latest version bound to PyTorch should just work
  • omegaconf, latest should just work
• Additional dependencies for ONNX examples:
  • onnx, latest should just work
  • onnxruntime, latest should just work
• Additional for TensorFlow examples:
  • tensorflow, latest should just work
  • tensorflow_hub, latest should just work

Please see the provided Colab for details for each example below. All examples are maintained to work with the latest major packaged versions of the installed libraries.

PyTorch

import torch
import zipfile
import torchaudio
from glob import glob

device = torch.device('cpu')  # gpu also works, but our models are fast enough for CPU
model, decoder, utils = torch.hub.load(repo_or_dir='snakers4/silero-models',
                                       model='silero_stt',
                                       language='en', # also available 'de', 'es'
                                       device=device)
(read_batch, split_into_batches,
 read_audio, prepare_model_input) = utils  # see function signature for details

# download a single file in any format compatible with TorchAudio
torch.hub.download_url_to_file('https://opus-codec.org/static/examples/samples/speech_orig.wav',
                               dst ='speech_orig.wav', progress=True)
test_files = glob('speech_orig.wav')
batches = split_into_batches(test_files, batch_size=10)
input = prepare_model_input(read_batch(batches[0]),
                            device=device)

output = model(input)
for example in output:
    print(decoder(example.cpu()))

ONNX

Our model will run anywhere that can import the ONNX model or that supports the ONNX runtime.

import onnx
import torch
import onnxruntime
from omegaconf import OmegaConf

language = 'en' # also available 'de', 'es'

# load provided utils
_, decoder, utils = torch.hub.load(repo_or_dir='snakers4/silero-models', model='silero_stt', language=language)
(read_batch, split_into_batches,
 read_audio, prepare_model_input) = utils

# see available models
torch.hub.download_url_to_file('https://raw.githubusercontent.com/snakers4/silero-models/master/models.yml', 'models.yml')
models = OmegaConf.load('models.yml')
available_languages = list(models.stt_models.keys())
assert language in available_languages

# load the actual ONNX model
torch.hub.download_url_to_file(models.stt_models.en.latest.onnx, 'model.onnx', progress=True)
onnx_model = onnx.load('model.onnx')
onnx.checker.check_model(onnx_model)
ort_session = onnxruntime.InferenceSession('model.onnx')

# download a single file in any format compatible with TorchAudio
torch.hub.download_url_to_file('https://opus-codec.org/static/examples/samples/speech_orig.wav', dst ='speech_orig.wav', progress=True)
test_files = ['speech_orig.wav']
batches = split_into_batches(test_files, batch_size=10)
input = prepare_model_input(read_batch(batches[0]))

# actual ONNX inference and decoding
onnx_input = input.detach().cpu().numpy()
ort_inputs = {'input': onnx_input}
ort_outs = ort_session.run(None, ort_inputs)
decoded = decoder(torch.Tensor(ort_outs[0])[0])
print(decoded)

TensorFlow

SavedModel example

import os
import torch
import subprocess
import tensorflow as tf
import tensorflow_hub as tf_hub
from omegaconf import OmegaConf

language = 'en' # also available 'de', 'es'

# load provided utils using torch.hub for brevity
_, decoder, utils = torch.hub.load(repo_or_dir='snakers4/silero-models', model='silero_stt', language=language)
(read_batch, split_into_batches,
 read_audio, prepare_model_input) = utils

# see available models
torch.hub.download_url_to_file('https://raw.githubusercontent.com/snakers4/silero-models/master/models.yml', 'models.yml')
models = OmegaConf.load('models.yml')
available_languages = list(models.stt_models.keys())
assert language in available_languages

# load the actual tf model
torch.hub.download_url_to_file(models.stt_models.en.latest.tf, 'tf_model.tar.gz')
subprocess.run('rm -rf tf_model && mkdir tf_model && tar xzfv tf_model.tar.gz -C tf_model',  shell=True, check=True)
tf_model = tf.saved_model.load('tf_model')

# download a single file in any format compatible with TorchAudio
torch.hub.download_url_to_file('https://opus-codec.org/static/examples/samples/speech_orig.wav', dst ='speech_orig.wav', progress=True)
test_files = ['speech_orig.wav']
batches = split_into_batches(test_files, batch_size=10)
input = prepare_model_input(read_batch(batches[0]))

# tf inference
res = tf_model.signatures["serving_default"](tf.constant(input.numpy()))['output_0']
print(decoder(torch.Tensor(res.numpy())[0]))

Text-To-Speech

Models and Speakers

All of the provided models are listed in the models.yml file. Any metadata and newer versions will be added there.

V4

V4 models support SSML. Also see Colab examples for main SSML tag usage.

ID	Speakers	Auto-stress	Language	SR	Colab
v4_ru	aidar, baya, kseniya, xenia, eugene, random	yes	ru (Russian)	8000, 24000, 48000
v4_cyrillic	b_ava, marat_tt, kalmyk_erdni...	no	cyrillic (Avar, Tatar, Kalmyk, ...)	8000, 24000, 48000
v4_ua	mykyta, random	no	ua (Ukrainian)	8000, 24000, 48000
v4_uz	dilnavoz	no	uz (Uzbek)	8000, 24000, 48000
v4_indic	hindi_male, hindi_female, ..., random	no	indic (Hindi, Telugu, ...)	8000, 24000, 48000

V3

V3 models support SSML. Also see Colab examples for main SSML tag usage.

ID	Speakers	Auto-stress	Language	SR	Colab
v3_en	en_0, en_1, ..., en_117, random	no	en (English)	8000, 24000, 48000
v3_en_indic	tamil_female, ..., assamese_male, random	no	en (English)	8000, 24000, 48000
v3_de	eva_k, ..., karlsson, random	no	de (German)	8000, 24000, 48000
v3_es	es_0, es_1, es_2, random	no	es (Spanish)	8000, 24000, 48000
v3_fr	fr_0, ..., fr_5, random	no	fr (French)	8000, 24000, 48000
v3_indic	hindi_male, hindi_female, ..., random	no	indic (Hindi, Telugu, ...)	8000, 24000, 48000

Dependencies

Basic dependencies for Colab examples:

• torch, 1.10+ for v3 models/ 2.0+ for v4 models;
• torchaudio, latest version bound to PyTorch should work (required only because models are hosted together with STT, not required for work);
• omegaconf, latest (can be removed as well, if you do not load all of the configs);

PyTorch

# V4
import torch

language = 'ru'
model_id = 'v4_ru'
sample_rate = 48000
speaker = 'xenia'
device = torch.device('cpu')

model, example_text = torch.hub.load(repo_or_dir='snakers4/silero-models',
                                     model='silero_tts',
                                     language=language,
                                     speaker=model_id)
model.to(device)  # gpu or cpu

audio = model.apply_tts(text=example_text,
                        speaker=speaker,
                        sample_rate=sample_rate)

Standalone Use

• Standalone usage only requires PyTorch 1.10+ and the Python Standard Library;
• Please see the detailed examples in Colab;

# V4
import os
import torch

device = torch.device('cpu')
torch.set_num_threads(4)
local_file = 'model.pt'

if not os.path.isfile(local_file):
    torch.hub.download_url_to_file('https://models.silero.ai/models/tts/ru/v4_ru.pt',
                                   local_file)  

model = torch.package.PackageImporter(local_file).load_pickle("tts_models", "model")
model.to(device)

example_text = 'В недрах тундры выдры в г+етрах т+ырят в вёдра ядра кедров.'
sample_rate = 48000
speaker='baya'

audio_paths = model.save_wav(text=example_text,
                             speaker=speaker,
                             sample_rate=sample_rate)

SSML

Check out our TTS Wiki page.

Cyrillic languages

Supported tokenset: !,-.:?iµöабвгдежзийклмнопрстуфхцчшщъыьэюяёђѓєіјњћќўѳғҕҗҙқҡңҥҫүұҳҷһӏӑӓӕӗәӝӟӥӧөӱӳӵӹ

Speaker_ID	Language	Gender
b_ava	Avar	F
b_bashkir	Bashkir	M
b_bulb	Bulgarian	M
b_bulc	Bulgarian	M
b_che	Chechen	M
b_cv	Chuvash	M
cv_ekaterina	Chuvash	F
b_myv	Erzya	M
b_kalmyk	Kalmyk	M
b_krc	Karachay-Balkar	M
kz_M1	Kazakh	M
kz_M2	Kazakh	M
kz_F3	Kazakh	F
kz_F1	Kazakh	F
kz_F2	Kazakh	F
b_kjh	Khakas	F
b_kpv	Komi-Ziryan	M
b_lez	Lezghian	M
b_mhr	Mari	F
b_mrj	Mari High	M
b_nog	Nogai	F
b_oss	Ossetic	M
b_ru	Russian	M
b_tat	Tatar	M
marat_tt	Tatar	M
b_tyv	Tuvinian	M
b_udm	Udmurt	M
b_uzb	Uzbek	M
b_sah	Yakut	M
kalmyk_erdni	Kalmyk	M
kalmyk_delghir	Kalmyk	F

Indic languages

Example

(!!!) All input sentences should be romanized to ISO format using aksharamukha. An example for hindi:

# V3
import torch
from aksharamukha import transliterate

# Loading model
model, example_text = torch.hub.load(repo_or_dir='snakers4/silero-models',
                                     model='silero_tts',
                                     language='indic',
                                     speaker='v4_indic')

orig_text = "प्रसिद्द कबीर अध्येता, पुरुषोत्तम अग्रवाल का यह शोध आलेख, उस रामानंद की खोज करता है"
roman_text = transliterate.process('Devanagari', 'ISO', orig_text)
print(roman_text)

audio = model.apply_tts(roman_text,
                        speaker='hindi_male')

Supported languages

Language	Speakers	Romanization function
hindi	hindi_female, hindi_male	transliterate.process('Devanagari', 'ISO', orig_text)
malayalam	malayalam_female, malayalam_male	transliterate.process('Malayalam', 'ISO', orig_text)
manipuri	manipuri_female	transliterate.process('Bengali', 'ISO', orig_text)
bengali	bengali_female, bengali_male	transliterate.process('Bengali', 'ISO', orig_text)
rajasthani	rajasthani_female, rajasthani_female	transliterate.process('Devanagari', 'ISO', orig_text)
tamil	tamil_female, tamil_male	transliterate.process('Tamil', 'ISO', orig_text, pre_options=['TamilTranscribe'])
telugu	telugu_female, telugu_male	transliterate.process('Telugu', 'ISO', orig_text)
gujarati	gujarati_female, gujarati_male	transliterate.process('Gujarati', 'ISO', orig_text)
kannada	kannada_female, kannada_male	transliterate.process('Kannada', 'ISO', orig_text)

Text-Enhancement

Languages	Quantization	Quality	Colab
'en', 'de', 'ru', 'es'	✔️	link

Dependencies

Basic dependencies for Colab examples:

• torch, 1.9+;
• pyyaml, but it's installed with torch itself

Standalone Use

• Standalone usage only requires PyTorch 1.9+ and the Python Standard Library;
• Please see the detailed examples in Colab;

import torch

model, example_texts, languages, punct, apply_te = torch.hub.load(repo_or_dir='snakers4/silero-models',
                                                                  model='silero_te')

input_text = input('Enter input text\n')
apply_te(input_text, lan='en')

Denoise

Denoise models attempt to reduce background noise along with various artefacts such as reverb, clipping, high/lowpass filters etc., while trying to preserve and/or enhance speech. They also attempt to enhance audio quality and increase sampling rate of the input up to 48kHz.

Models

All of the provided models are listed in the models.yml file.

Model	JIT	Real Input SR	Input SR	Output SR	Colab
small_slow	✔️	8000, 16000, 24000, 44100, 48000	24000	48000
large_fast	✔️	8000, 16000, 24000, 44100, 48000	24000	48000
small_fast	✔️	8000, 16000, 24000, 44100, 48000	24000	48000

Dependencies

Basic dependencies for Colab examples:

• torch, 2.0+;
• torchaudio, latest version bound to PyTorch should work;
• omegaconf, latest (can be removed as well, if you do not load all of the configs).

PyTorch

import torch

name = 'small_slow'
device = torch.device('cpu')
model, samples, utils = torch.hub.load(
  repo_or_dir='snakers4/silero-models',
  model='silero_denoise',
  name=name,
  device=device)
(read_audio, save_audio, denoise) = utils

i = 0
torch.hub.download_url_to_file(
  samples[i],
  dst=f'sample{i}.wav',
  progress=True
)
audio_path = f'sample{i}.wav'
audio = read_audio(audio_path).to(device)
output = model(audio)
save_audio(f'result{i}.wav', output.squeeze(1).cpu())

i = 1
torch.hub.download_url_to_file(
  samples[i],
  dst=f'sample{i}.wav',
  progress=True
)
output, sr = denoise(model, f'sample{i}.wav', f'result{i}.wav', device='cpu')

Standalone Use

import os
import torch

device = torch.device('cpu')
torch.set_num_threads(4)
local_file = 'model.pt'

if not os.path.isfile(local_file):
    torch.hub.download_url_to_file('https://models.silero.ai/denoise_models/sns_latest.jit',
                                   local_file)  

model = torch.jit.load(local_file)
torch._C._jit_set_profiling_mode(False) 
torch.set_grad_enabled(False)
model.to(device)

a = torch.rand((1, 48000))
a = a.to(device)
out = model(a)

FAQ

Wiki

Also check out our wiki.

Performance and Quality

Please refer to these wiki sections:

• Quality Benchmarks
• Performance Benchmarks

Adding new Languages

Please refer here.

Contact

Get in Touch

Try our models, create an issue, join our chat, email us, and read the latest news.

Commercial Inquiries

Please refer to our wiki and the Licensing and Tiers page for relevant information, and email us.

Citations

@misc{Silero Models,
  author = {Silero Team},
  title = {Silero Models: pre-trained enterprise-grade STT / TTS models and benchmarks},
  year = {2021},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/snakers4/silero-models}},
  commit = {insert_some_commit_here},
  email = {hello@silero.ai}
}

Further reading

English

• STT:

  • Towards an Imagenet Moment For Speech-To-Text - link
  • A Speech-To-Text Practitioners Criticisms of Industry and Academia - link
  • Modern Google-level STT Models Released - link

• TTS:

  • Multilingual Text-to-Speech Models for Indic Languages - link
  • Our new public speech synthesis in super-high quality, 10x faster and more stable - link
  • High-Quality Text-to-Speech Made Accessible, Simple and Fast - link

• VAD:

  • One Voice Detector to Rule Them All - link
  • Modern Portable Voice Activity Detector Released - link

• Text Enhancement:

  • We have published a model for text repunctuation and recapitalization for four languages - link

Chinese

• STT:
  • 迈向语音识别领域的 ImageNet 时刻 - link
  • 语音领域学术界和工业界的七宗罪 - link

Russian

• STT

  • OpenAI решили распознавание речи! Разбираемся так ли это … - link
  • Наши сервисы для бесплатного распознавания речи стали лучше и удобнее - link
  • Telegram-бот Silero бесплатно переводит речь в текст - link
  • Бесплатное распознавание речи для всех желающих - link
  • Последние обновления моделей распознавания речи из Silero Models - link
  • Сжимаем трансформеры: простые, универсальные и прикладные способы cделать их компактными и быстрыми - link
  • Ультимативное сравнение систем распознавания речи: Ashmanov, Google, Sber, Silero, Tinkoff, Yandex - link
  • Мы опубликовали современные STT модели сравнимые по качеству с Google - link
  • Понижаем барьеры на вход в распознавание речи - link
  • Огромный открытый датасет русской речи версия 1.0 - link
  • Насколько Быстрой Можно Сделать Систему STT? - link
  • Наша система Speech-To-Text - link
  • Speech-To-Text - link

• TTS:

  • Теперь наш синтез также доступен в виде бота в Телеграме - link
  • Может ли синтез речи обмануть систему биометрической идентификации? - link
  • Теперь наш синтез на 20 языках - link
  • Теперь наш публичный синтез в супер-высоком качестве, в 10 раз быстрее и без детских болячек - link
  • Синтезируем голос бабушки, дедушки и Ленина + новости нашего публичного синтеза - link
  • Мы сделали наш публичный синтез речи еще лучше - link
  • Мы Опубликовали Качественный, Простой, Доступный и Быстрый Синтез Речи - link

• VAD:

  • Наш публичный детектор голоса стал лучше - link
  • А ты используешь VAD? Что это такое и зачем он нужен - link
  • Модели для Детекции Речи, Чисел и Распознавания Языков - link
  • Мы опубликовали современный Voice Activity Detector и не только -link

• Text Enhancement:

  • Восстановление знаков пунктуации и заглавных букв — теперь и на длинных текстах - link
  • Мы опубликовали модель, расставляющую знаки препинания и заглавные буквы в тексте на четырех языках - link

Donations

Please use the "sponsor" button.