silero_tts_standalone is a simple script which can be used to TTS large text with Silero TTS models locally (do txt -> wav conversion).
By default, script is configured for Russian texts, but it can be reconfigured for any language supported by Silero models.
In order to work with non-Russian texts you should comment out spell_digits() function and its call in preprocess_text(), or (better) rewrite it with a module supporting your language. You also should translate replacement strings in preprocess_text() according to your text language.
The script was created to operate with large texts (over 1 MiB) but can handle small texts too.
It provides the following features:
- Basic text preprocessing (replace unsupported by model characters to supported, replace digits like 11 with "одиннадцать" to TTS them, limit line length according to punctuation)
- WAV file size limiting (WAV format is limited to 4 GiB file size) according to sentences (no awkward mid-word splits)
- Verbose run-time output with runtime estimation, full TTS size and length estimation and timestamps for each TTSed line
Usage: ./tts.py text.txt
The script was tested only with UTF-8 texts.
During runtime, it will output the following lines:
3/341 0:00:05/0:17:04 469/96065 chars 2/522 MiB 0:00:27/1:32:10 TTS 0:00:27@part0 0.5% : В ответ
- 3 - current line number
- 341 - total lines count
- 0:00:05 - elapsed time
- 0:17:04 - estimated time
- 469 - processed characters
- 96065 - total characters
- 2 - WAV size already written to output files (total)
- 522 - estimated WAV sizes (total)
- 0:00:27 - line timestamp (total)
- 1:32:10 - estimated length of all files
- 0:00:27 - line timestamp in current WAV file
- part0 - current WAV file number
- 0.5% - progress
- В ответ - processed string
Estimations may be inaccurate right after start, but after ~1 minute it will be more or less reliable.
Script will output the following files:
- ${INPUT_FILENAME}_preprocessed.txt - preprocessed text (it will be TTSed)
- ${INPUT_FILENAME}_0.wav
- ${INPUT_FILENAME}_1.wav
- ... - TTS result
Requirements:
- Python 3.10.7+ (may work on earlier versions, but not tested)
- pytorch
- numpy
- num2t4ru (for spell_digits())
