New STT provider: gigastt — local, CPU-only, best-in-class Russian ASR

[ekhodzitsky]

Hi!

I saw "Add more provider integrations (STT, TTS)" on your roadmap, so here's a candidate.

I'm building gigastt — a self-hosted speech-to-text server built on GigaAM v3 (SberDevices), MIT licensed: https://github.com/ekhodzitsky/gigastt

Quick facts, all measured on CPU (M1) through a single cross-engine harness, 1000 samples per domain:

• Russian WER: 3.55% clean read vs 15.53% for faster-whisper Large v3; phone calls 18.50% vs 24.93%; YouTube 10.91% vs 15.45%
• RTF ~0.10 on plain CPU — faster-whisper is slower than real-time on CPU
• ~225 MB INT8 model, cold start under 1s, ~400 MB RSS per session
• REST /v1/transcribe plus genuine WebSocket streaming (PCM16 in, incremental partials out, TTFP ~780 ms)
• Docker image out of the box, no API keys, fully offline
• Punctuation and ITN are restored on top of the core model by a bundled RuPunct model

Full methodology and honest caveats: https://github.com/ekhodzitsky/gigastt/blob/main/docs/benchmarks.md

To be upfront about scope: gigastt is Russian-first (plus GigaAM Multilingual heads covering ru/en/kk/ky/uz — 4.63% WER on LibriSpeech test-clean for English). So it's not a wholesale Whisper replacement — it's a provider that gives your Russian-speaking users a dramatic accuracy jump while removing the Deepgram dependency for those calls.

If you're open to it, I'd be happy to write the provider integration myself and send a PR.