Russian speech recognition for Voice input — GigaAM-v3 optional bundle (working prototype, MIT, no new native dependency)

[almaziphone]

TL;DR

I built a community bundle that adds Russian to the experimental Voice input capability:
GigaAM-v3 e2e CTC INT8 (MIT, SaluteDevices/Sber) running on the already shipped
sherpa-onnx-node runtime. It registers a second provider into ctx.speechToText, so the
microphone UI, preparation cards, cancellation, settings persistence and transcript insertion work
unchanged.

• Repository (public, MIT, dsh-plugin topic): https://github.com/almaziphone/dsh-speech-to-text-ru
• Install on DSH Desktop:

  bash -c "$(curl -fsSL https://raw.githubusercontent.com/almaziphone/dsh-speech-to-text-ru/main/scripts/install.sh)" install.sh

  or manually: enable Plugins → Voice input, then
  dsh plugin --profile desktop add github:almaziphone/dsh-speech-to-text-ru, refresh the window.

Why this may be worth shipping

• Voice input currently has exactly one local recognizer — SenseVoiceSmall, whose hints are
  auto, zh, en, yue, ja, ko. There is no cloud provider in the experimental composition, so
  Russian-speaking users cannot use voice input at all.
• GigaAM-v3 is the SOTA open Russian ASR family; its e2e variant already emits punctuation,
  casing and number normalization (e.g. «в 15:30»), which is exactly what dictation needs.
• No new native dependency: sherpa-onnx-node is already a dependency of
  packages/experimental/speech-to-text-sensevoice. The model is a pinned, on-demand download
  (321.7 MB model + 2 KB tokens + 1.8 MB Silero VAD), verified by size and SHA-256 exactly like
  SenseVoice assets.
• Provider architecture mirrors the shipped one: host-local, languages ru/auto, preparation
  state machine (check → download → verify → load), bounded queue, managed worker process over an
  authenticated loopback listener, Silero VAD segmentation, idle reclamation.
• Licenses: GigaAM MIT, sherpa-onnx Apache-2.0, Silero VAD MIT. Weights are downloaded, never
  redistributed; no new licensing burden for the installation.

Measured on the app's own runtime (sherpa-onnx-node 1.13.8, macOS arm64, reference set)

	GigaAM-v3 e2e CTC INT8 (proposed)	Whisper Small INT8 (fallback engine in the package)
WER	3.03 % (0 lexical errors)	13.64 %
Exact normalized matches	4 / 5 clips	2 / 5
RTF @ 2 threads	0.017 (8.4 s clip → ~0.17 s)	0.257
Peak RSS	564 MiB	1.40 GiB
Model load	0.36 s	0.61 s
Download	321.7 MB + 1.8 MB VAD	375.5 MB + 1.8 MB VAD

Two native quirks are guarded in the worker and covered by tests: sherpa-onnx aborts on decodes
shorter than 160 samples, and the literal language "auto" aborts the Whisper engine (only ""
means auto-detect).

What a PR would contain (when external PRs are accepted)

CONTRIBUTING.md says external pull requests are not accepted at the moment, so I'm posting this
here as the suggested channel instead of opening one. The community package is the working
prototype; porting it to the repository conventions is mechanical and I'm happy to do it:

1. New packages/experimental/speech-to-text-gigaam/ mirroring speech-to-text-sensevoice —
   TypeScript source (src/index.ts provider, src/worker.ts), pinned runtime/assets.json,
   locale/{en,zh}.json, README.{md,zh.md} + README.i18n.yaml, tests/, tsconfig.json,
   tsdown.config.ts.
2. packages/experimental/voice-input-bundle/cordis.patch.yml: insert the provider row
   (id: speech-to-text-gigaam, dataRoot: !!js dshHomePath('speech-to-text', 'gigaam')) and keep
   sensevoice-local as the shipped default — users pick the recognizer in Settings → Voice input.
3. Release manifests (apps/desktop/package.json and the runtime dependency list) plus
   scripts/optional-bundles.spec.ts / installed-product-isolation expectations.
4. Docs: docs/subsystems/voice-input.md (+ .zh.md, .i18n.yaml), the experimental package
   table, and THIRD_PARTY_NOTICES.md (GigaAM attribution).
5. Model-free tests: language mapping, config normalization, the <160-sample guard, WAV
   validation, worker transport — CI-friendly, no downloads.

Open questions: extend the existing voice-input bundle (my preference) or ship a separate Russian
bundle; whether one more CPU-only language model fits the optional-bundle budget; Chinese docs
review (I can draft EN and adapt to your conventions).

Proof

Installed from GitHub into a clean DSH home, the host recomposed with the provider registered
(gigaam-ru, languages ru/auto), preparation reused a verified cache, and a Russian recording
was transcribed through the real speechToText service in ~1.8 s wall time with punctuation and
number normalization. Benchmarks and methodology: docs/asr-validation.md in the repo;
implementation-ready integration checklist: docs/upstream.md.

AI declaration

The community implementation and this proposal were produced with an AI coding agent under my
direction; all measurements were run on my machine against the app's bundled runtime. If the work
is ported into the repository, I will fill the AI-declaration table per
.agents/skills/dsh-pr-description and handle review feedback.

Thanks for reading! — @almaziphone