The developer of the Telegram bot @chisto_text_bot has implemented a free, ad-free transcription service for voice messages, video notes, and audio files. The system operates on a single CPU core with up to 1.5 GB of RAM, avoiding the need for a GPU.
The bot utilizes the open-source GigaAM v3 model, which achieves a 7.9% word error rate on conversational speech. It processes one minute of Russian audio in approximately 2 seconds, significantly outperforming Whisper small, which requires 14–30 seconds per minute of audio and has a higher error rate of 14–18%.
To handle long files up to 3 hours or 2 GB, the system uses a custom energy-based segmentation method instead of a dedicated VAD model. It processes audio in 20–30 second chunks and employs a round-robin scheduling algorithm to ensure that short voice messages are not blocked by long-running transcription tasks.
Privacy is maintained by processing all files locally on the server; data is deleted immediately after transcription. The bot uses a local instance of the Telegram Bot API to handle large files, and systemd is used to restrict resource usage and ensure security.
