Examples

Jupyter notebooks in the notebooks/ directory of the repository. Each opens in Google Colab with a button and installs stt-ru itself; the outputs of all cells are stored in the files, so the notebooks can be read without running, and the recordings can be played in the embedded players. The notebooks download their recordings and references themselves. The notebooks are in Russian.

Quick start

01_quickstart.ipynb - a stereo phone call from reading the file to the result: load_audio and Transcriber, segments by speaker in a table, a waveform from Audio.peaks with the detected speech segments, every output format via save and render, mixing the channels down (split_channels=False), transcribing a single channel from memory via transcribe_audio, and the same work with the stt-ru transcribe command.

Call analytics

02_call_analytics.ipynb - what the transcription of a stereo recording tells about a conversation: a timeline of turns by role, speech and silence time, pace in words per minute, response time, backchannels and interruptions, searching for key phrases in the normalized text with regular expressions, transcribing a folder of calls with the command and a --summary file, and a summary table by call.

Long recording and subtitles

03_long_recording.ipynb - the first chapter of the audiobook "Dubrovsky" (17 minutes, MP3, LibriVox): reading through ffmpeg, reading and recognition time separately, speed relative to real time, SRT and VTT subtitles, the distribution of segment durations, and tuning the speech detection parameters min_silence_duration_ms and max_speech_duration_s for subtitles.

Evaluating models on your own references

04_model_evaluation.ipynb - how to choose a model for your recordings with the stt-ru evaluate command: a CSV manifest of 100 FLEURS phrases, WER, CER and speed of four models, the detailed JSON report, the effect of numbers written as digits or words, text normalization, the recordings with the most errors, int8 against original weights from Python, and the format of your own manifest. The notebook shows how to use the evaluation; choose a model on your own recordings.

Web service and REST API from Python

05_web_service.ipynb - stt-ru serve in the background with an access token and working with it through httpx: queueing a file and waiting for the result, the result in different formats, the waveform and the recording with seeking, error responses, deleting a job, a client for a batch of files, and running the same service in Docker.

Running locally

uv sync --all-extras --group notebooks
uv run jupyter lab notebooks/

The stt-ru commands in the notebooks run in %%bash cells: bash is available on Linux, macOS and Colab; on Windows the bash from Git for Windows must be in PATH.

make notebooks runs all notebooks, writes the outputs to the files and removes the execution metadata; the notebooks.yml workflow runs the notebooks on a schedule and on demand.