The developer integrated GigaChat and GigaAM models to automate stress marking and quality control for Russian text-to-speech (TTS) systems. By processing a bank of 7,789 words through GigaChat-2-Pro, the system generated a lexicon covering 7,462 words, expanding the manual dictionary from 51 to 101 entries.
The process revealed that large-batch processing leads to inconsistent results. The final workflow requires verifying candidates in small batches and manual review of discrepancies. The developer noted that GigaChat-2-Lite proved ineffective for this task, while the Pro version provided the necessary accuracy.
For quality control, GigaAM v3 was implemented to transcribe generated audio. While efficient for clean audio, it struggles with radio-effect processing, where Whisper remains the preferred tool. Due to processing limits, audio files are segmented into 30-second windows for analysis.
The integration is managed via custom scripts, including stress_suggest.py for dictionary generation and qa_recent.py for verification. The developer successfully used GigaAM to transcribe a six-minute voice recording with high accuracy on clean audio.





