Marude

Marusia demo interface - an http api client which allows to automatically convert russian short texts into speech using vk cloud.

Installation

To install system-wise dependencies:

sudo apt-get update && sudo apt-get install build-essential libtag1-dev ffmpeg

To create a conda environment and install dependencies use the following command:

conda env create -f environment.yml
python -c 'import nltk; nltk.download("punkt")'

Then activate created environment:

conda activate marude

Usage

Anecdotes

There is an auxiliary script fetch.sh for generating baneks dataset, which should be used invoked this:

./fetch.sh 05.11.2025

This script accepts version name which may be an arbitrary string.

Then to transform the generated dataset to speech:

python -m rr handle-aneks -s assets/baneks/05.11.2025/default.tsv -d assets/baneks-speech -e salute -rk

Where assets/baneks-speech should contain previously generated audio files to avoid making duplicates.

Then to copy new files to a separate folder use a command like:

ls -alth | grep Nov | grep mp3 | awk '{print $9}' | xargs -I {} mv {} ../baneks-speech-new/

And to generate an archive run the following command:

tar -cJvf ../../baneks-speech/speech/040001-040720.tar.xz -C /home/zeio/raconteur/assets/baneks-speech-new/ .

Etc

After the environment is set up, the app can be used from the command line:

python -m marude tts 'Привет, мир' -m pavel -p message.mp3

The provided text (which must be 1024 characters long or shorter) will be converted into speech and saved as an audiofile message.mp3. By default the file is saved at assets/message.mp3.

The module can be used programmatically as well. First, install the system-wise dependencies:

sudo apt-get update && sudo apt-get install ffmpeg libtag1-dev

Then, install the module through pip:

pip install marude

Then run your script, which may look like this (see example):

from tasty import pipe
from marude import CloudVoiceClient, Voice


if __name__ == '__main__':

    client = CloudVoiceClient(Voice.MARIA)

    with open('message-1.mp3', 'wb') as file:
        _ = 'Съешь еще этих мягких французских булок' | pipe | client.tts | pipe | file.write

    with open('message-2.mp3', 'wb') as file:
        _ = 'да выпей чаю' | pipe | client.tts | pipe | file.write

Automatic speech recognition

Automatic speech recognition (ASR) pipeline is implemented as a pair of services:

1. producer - splits input mp3 file into chunks with given max-length on silence and converts audio to wav format;
2. consumer - as soon as producer completes converting the next chunk it loads the file and sends it to a remote service for recognition. Then it appends the recognized text to the output file.

Example of command for running the pipeline:

python -m marude asr assets/real-colonel.mp3

Example of log looks like this (the whole pipeline was completed in 28 seconds):

Finished converting assets/real-colonel.mp3 to .wav which is saved as assets/real-colonel-converted.wav. Audio duration is 1502.856 seconds
Started segmenting assets/real-colonel-converted.wav
Finished segmenting assets/real-colonel-converted.wav. There are 413 segments
Final number of segments is 15
Started saving segments on disk
Finished saving 01/15 segments. The latest segment was saved as assets/real-colonel/00.wav
Finished saving 02/15 segments. The latest segment was saved as assets/real-colonel/01.wav
Finished saving 03/15 segments. The latest segment was saved as assets/real-colonel/02.wav
Finished saving 04/15 segments. The latest segment was saved as assets/real-colonel/03.wav
Finished saving 05/15 segments. The latest segment was saved as assets/real-colonel/04.wav
Finished saving 06/15 segments. The latest segment was saved as assets/real-colonel/05.wav
Finished saving 07/15 segments. The latest segment was saved as assets/real-colonel/06.wav
Finished saving 08/15 segments. The latest segment was saved as assets/real-colonel/07.wav
Finished saving 09/15 segments. The latest segment was saved as assets/real-colonel/08.wav
Finished saving 10/15 segments. The latest segment was saved as assets/real-colonel/09.wav
Finished saving 11/15 segments. The latest segment was saved as assets/real-colonel/10.wav
Finished saving 12/15 segments. The latest segment was saved as assets/real-colonel/11.wav
Recognized text from file assets/real-colonel/00.wav
Finished saving 13/15 segments. The latest segment was saved as assets/real-colonel/12.wav
Finished saving 14/15 segments. The latest segment was saved as assets/real-colonel/13.wav
Finished saving 15/15 segments. The latest segment was saved as assets/real-colonel/14.wav
Recognized text from file assets/real-colonel/01.wav
Recognized text from file assets/real-colonel/02.wav
Recognized text from file assets/real-colonel/03.wav
Recognized text from file assets/real-colonel/04.wav
Recognized text from file assets/real-colonel/05.wav
Recognized text from file assets/real-colonel/06.wav
Recognized text from file assets/real-colonel/07.wav
Recognized text from file assets/real-colonel/08.wav
Recognized text from file assets/real-colonel/09.wav
Recognized text from file assets/real-colonel/10.wav
Recognized text from file assets/real-colonel/11.wav
Recognized text from file assets/real-colonel/12.wav
Recognized text from file assets/real-colonel/13.wav
Recognized text from file assets/real-colonel/14.wav

See details of this example here. There is also another example with 3-hour recording, which was handled in ~ 12 minutes.