Benchmarking Russian speech recognition on CPU: T-one, GigaAM-v3 and cloud models

Description and use

Benchmarked Russian speech recognition on CPU: T‑one (streaming, 0.31 RTF), GigaAM‑v3 (offline, 9.4% distance to consensus, 13 s), and Gemini cloud models (7–8.6% distance, 12–61 s).

Problem

Need a voice agent that understands Russian speech over the phone and answers out loud, requiring reliable on‑CPU speech recognition.

Outcome

T‑one runs real‑time on a single CPU thread with 0.31 RTF; GigaAM‑v3 achieves 9.4% distance to consensus and 13 s median time on CPU; Gemini cloud models achieve 7.0–8.6% distance but are slower and more expensive.

Workflow

• 1. Gather seven personal voice‑message recordings (7–120 s).
• 2. Run each recording through T‑one (greedy decoder, 1 thread) and record time, RTF, CPU time.
• 3. Measure latency per phrase and adjust MIN_SILENCE_DURATION to reduce lag.
• 4. Run recordings through GigaAM‑v3 (offline, int8 encoder) and record median time per file.
• 5. Compare with Gemini cloud models (audio input) and compute distance to consensus and median time.
• 6. Document installation challenges (kenlm build, decoder selection).
• 7. Summarize results in tables and screenshots.
• 8. Store scripts and license checks in ru‑speech‑bench repo.

Evidence

On a 53‑second clip, greedy decoder, best of four runs:

Source

Disagreement against GigaAM as the reference: 20‑22% on a clean clip, 41.5‑46.8% on a quiet recording of two people.

Source

On my seven recordings on the same VM, four threads, under parallel load: median 12.8 s per file for lengths of 7 to 130 s, range 3.8‑26.6 s.

Source

Gemini‑3.6‑flash did best on Latin terms, but results are unstable from version to version.

Source