The author developed a two-stage transcription system for meetings. It uses the T-one model for real-time draft subtitles and the GigaAM-v3 model to generate a high-quality final document after the call.
Testing was conducted on a server with a 6-vCPU Xeon E5-2680 v2 processor and 11 GB of RAM. The system handles two simultaneous meetings, delivering phrases with a latency of 1.08 to 1.2 seconds.
The final text generation process, which corrects 17% to 26% of the draft, runs with low priority. This ensures that real-time subtitles remain unaffected by the background processing load.
The author found that a secondary model for correcting terms was ineffective, often increasing the Word Error Rate (WER). Instead, a custom personal dictionary approach proved more efficient for fixing terminology.
The system successfully handles Russian-language meetings with English terms. While the final document quality is comparable to cloud solutions like Deepgram, the author notes that performance depends on CPU architecture.
