Repository navigation
Mluva 2.5.0 adds three cloud speech providers chosen by a Czech and English benchmark. Microsoft MAI-Transcribe-2 is listed first as Suggested: in live dictation it halves Scribe v2's Czech word error rate (5.3% against 10.8%) and returns the final text fastest after Stop. xAI Grok Voice Transcribe 2.0 and Mistral Voxtral 3 join ElevenLabs Scribe, the local models and a compatible endpoint in Settings → Providers. Existing installations keep the provider they saved.
- Pick Realtime or Batch for every cloud provider, with its price per audio hour beside the choice. Batch sends the finished recording at Stop; MAI costs $0.54 per audio hour live and $0.10 in batch (introductory price until 31 December 2026).
- Read Why these providers on the Providers page: word error rate, delay and price for every route on 62 public clips, and why Alibaba Fun-ASR is not offered. Voxtral Small is the most accurate batch route in English; Scribe remains strongest on Czech technical terms.
- Keep each cloud key in its own desktop-keyring entry. The Azure Speech region is a setting, never a secret.
- When a realtime stream fails, leaves words unfinished, or returns no text for a recording of a second or more, Mluva recognizes the retained recording in batch and says so.
- Settings saved by this release stay readable by 2.4.0 until you choose a new provider, mode or region.
Download mluva-2.5.0-omarchy-x86_64.tar.gz and SHA256SUMS, verify with sha256sum --ignore-missing --check SHA256SUMS, extract the archive, quit Mluva, and run bash install.sh inside it. The source archive is also attached. Existing settings, History, retained audio, models and keyring entries are preserved.
Verification: full native gate (403 passed, 65 explicitly ignored integration tests), strict Clippy, formatting, ShellCheck, shortcut portal checks (21 workflows / 59 states), hosted package checks, archive inventory validation (594 files, 2 links), and isolated install/startup smokes for the previous settings and each new provider. The new routes were exercised against local protocol peers, a full capture session and the isolated desktop runner, and once each against the live services with public clips. Release evidence · Provider evidence. Hosted CI was not run; it is manual-only.
The new routes are Experimental. Physical microphone and Hyprland acceptance remain unverified. The macOS app still dictates with ElevenLabs only; its two mirrored settings were not built here. MAI is a Microsoft preview without an SLA, the Azure free tier allows one live stream at a time, and Mistral's entry tier rate-limits bursts. The optional Tensaku editor and private phone bridge artifacts remain in v2.3.0.
Changes: #126 and the 2.5.0 release metadata in #128. Fixes S27-794.