Speech to text transcribation services

Cloning the repository

git clone https://github.com/format37/stt.git
cd stt

Kaldi Vosk speech to text transcribation GPU docker server

Installation

cd vosk
pip install -r requirements.txt

Download the model

Choose your language:

• English

wget https://alphacephei.com/vosk/models/vosk-model-en-us-0.22.zip
unzip vosk-model-en-us-0.22.zip
mv vosk-model-en-us-0.22 model
rm vosk-model-en-us-0.22.zip

• Russian

wget https://alphacephei.com/vosk/models/vosk-model-ru-0.10.zip
unzip vosk-model-ru-0.10.zip
mv vosk-model-ru-0.10 model
rm vosk-model-ru-0.10.zip

• Another languages:
  List of models

Build

docker-compose up -d

Using

python asr-test.py en.wav

Expected output:

# 1.0 1.26 1.59 some
# 1.0 1.59 2.01 people
# 1.0 2.04 2.220728 have
# 0.965297 2.220728 2.64 already
# 1.0 3.12 3.75 committed
# 1.0 3.81 3.93 to
# 1.0 3.93 4.44 memory
# 1.0 4.86 5.13 but
# 1.0 5.16 5.34 if
# 1.0 5.34 5.43 you
# 1.0 5.43 5.91 haven't
# 1.0 5.94 6.09 you
# 0.74926 6.09 6.39 shouldn't
# 1.0 6.39 6.81 before
# 1.0 6.81 7.05 your
# 1.0 7.14 7.77 interview
=== middle confidence: 0.9821598125 

# 1.0 8.91 8.97 the
# 1.0 8.97 9.27 table
# 1.0 9.27 9.66 comes
# 1.0 9.66 9.81 in
# 1.0 9.81 10.2 handy
# 1.0 10.23 10.77 often
# 1.0 11.22 11.4 in
# 1.0 11.4 12.0 scalability
# 1.0 12.0 12.51 question
# 0.967827 12.51 12.66 in
# 0.795548 12.66 13.05 computer
# 1.0 13.08 13.26 how
# 1.0 13.26 13.59 much
# 1.0 13.59 14.01 space
# 1.0 14.1 14.19 a
# 1.0 14.19 14.49 set
# 1.0 14.49 14.64 of
# 1.0 14.64 15.12 data
# 1.0 15.36 15.54 will
# 1.0 15.54 15.84 take
# 1.0 15.84 16.14 up
=== middle confidence: 0.9887321428571428 

["some people have already committed to memory but if you haven't you shouldn't before your interview", 'the table comes in handy often in scalability question in computer how much space a set of data will take up']

Troubles

It should be a Mono audio file, with Sample rate, the same as set in docker-compose.yml
To prepare audio file we can use this command:

ffmpeg -i audio.wav -ac 1 -ar 16000 audio_prepared.wav

Thanks to

This container based on Sergey Korol's repository and his docker image

Yandex speech kit direct request

Asynchronous recognition:

• Maximum recording duration: 4 hours.
• Maximum file size: 1 GB.

Supported languages:

• ru-RU (default): Russian.
• kk-KK: Kazakh.

Requirements

• Yandex cloud paid account
• Uploaded to blob storage audio file

Example:

direct.ipynb