dsh-speech-to-text-ru

Local Russian speech recognition for DeepSeek Harness voice input. Adds a second, fully offline recognizer — GigaAM-v3 e2e CTC int8 — next to the shipped SenseVoice provider, which understands only Chinese, English, Cantonese, Japanese and Korean. Audio never leaves the machine.

┌───────────────┐     ┌──────────────────────┐     ┌───────────────────────────┐
│  Microphone   │ ──► │  DSH speechToText    │ ──► │  gigaam-ru (this bundle)  │
│  (voice UI)   │     │  service (Host)      │     │  GigaAM-v3 CTC int8, CPU  │
└───────────────┘     └──────────────────────┘     └───────────────────────────┘
                        pinned model download + SHA-256 verify + managed worker

Why

DSH's built-in Voice input plugin transcribes with SenseVoiceSmall, which has no Russian support. This bundle registers another speechToText provider backed by GigaAM-v3 (Sber's Russian ASR foundation model, MIT) running on the sherpa-onnx native runtime that DSH already ships. No cloud, no API keys, no Python, no extra services.

Highlights

Model	GigaAM-v3 e2e CTC INT8 (punctuation, casing and number normalization included)
Quality	3.0 % WER on the reference Russian set, 0 lexical errors (only inverse text normalization differences)
Speed	RTF 0.017 @ 2 threads — an 8.4 s phrase decodes in ~0.17 s, model load ~0.36 s
Memory	peak RSS ~564 MiB; the worker is stopped after idleTimeoutMs
Download	319.9 MB model + 2 KB tokens + 1.8 MB Silero VAD, verified by size and SHA-256
Languages	ru (native), auto; multilingual Whisper fallback engine adds en and 99 more
Privacy	CPU inference on the Host; audio never leaves the machine
Platforms	macOS arm64/x64, Linux x64/arm64, Windows x64 (whatever DSH Desktop supports)

Requirements

• DSH Desktop ≥ 0.2.1-rc.5 (or a CLI/Web profile composed from dsh-base + dsh-web-app).
• ~330 MB disk for the model cache and ~0.6 GB RAM while the recognizer is loaded.
• Network access for the first model download (Hugging Face, with hf-mirror.com fallback). Fully offline machines can pre-seed the cache — see Offline install.

Install

Order matters. Enable the built-in Voice input first, then add this bundle. DSH applies bundle patches in dsh.profile.bundles order, so the Russian bundle has to sit after the voice-input bundle to make itself the default recognizer. The one-command installer below does this for you; with the manual route, pick GigaAM-v3 · RU (int8) once in Settings if the order was the other way round (DSH persists the selection), or append the small patch block that scripts/install.sh writes.

One command (recommended)

# one line, no clone needed:
bash -c "$(curl -fsSL https://raw.githubusercontent.com/almaziphone/dsh-speech-to-text-ru/main/scripts/install.sh)" install.sh

# or from a clone (add `--spec .` to install the checkout itself instead of GitHub):
git clone https://github.com/almaziphone/dsh-speech-to-text-ru
cd dsh-speech-to-text-ru
./scripts/install.sh

The script uses the dsh CLI that ships with DSH Desktop, backs up the profile, enables Voice input, installs/selects this bundle (dsh plugin add, so pnpm pulls the pinned sherpa-onnx-node@1.13.8 runtime for your platform) and pins defaultProvider: gigaam-ru, language: ru in the profile patch — so later toggling cannot change the default.

Manual

1. In DSH open Plugins and switch Voice input on — that adds the microphone UI, the speechToText service and its settings.

2. Add the bundle (the dsh CLI is /Applications/DeepSeek Harness.app/Contents/Resources/runtime/cli/bin/dsh on macOS; add it to PATH or call it by full path):

   dsh plugin --profile desktop add github:almaziphone/dsh-speech-to-text-ru
   

3. Refresh the app window (Cmd+R / Ctrl+R). Open Settings → Voice input: the recognizer should be GigaAM-v3 · RU (int8) with language ru. If it is not (for example you enabled Voice input after installing the bundle), select it once — DSH stores the choice per profile.

Other sources:

dsh plugin --profile desktop add ./dsh-speech-to-text-ru-0.1.0.tgz   # release tarball
dsh plugin --profile desktop add dsh-speech-to-text-ru               # npm, once published
dsh plugin --profile desktop add /path/to/cloned/checkout            # local checkout (link:)

Distribution

The package is a normal DSH bundle, so it can be distributed in several ways:

Source	Command
GitHub (recommended)	dsh plugin --profile desktop add github:almaziphone/dsh-speech-to-text-ru
Release tarball	dsh plugin --profile desktop add ./dsh-speech-to-text-ru-0.1.0.tgz
npm registry	dsh plugin --profile desktop add dsh-speech-to-text-ru
Local checkout	dsh plugin --profile desktop add /path/to/checkout

Model weights are not bundled: the published package is ~30 KB, while the pinned GigaAM-v3 INT8 weights (319.9 MB) and Silero VAD (1.8 MB) — ~322 MB together — are downloaded once on the target machine at first preparation and verified against the pinned SHA-256 hashes. Offline machines can copy ~/.dsh/speech-to-text/gigaam-ru/ instead, or use scripts/install-from-checkout.sh on a clone.

Maintainers publish with:

npm publish --access public   # from this repository (package.json is the bundle manifest)
gh release create v0.1.0 --generate-notes dsh-speech-to-text-ru-0.1.0.tgz

First use

The model is not bundled with the package. On first preparation DSH downloads the pinned assets from Hugging Face (revision pinned, size + SHA-256 verified while streaming) into ~/.dsh/speech-to-text/gigaam-ru/:

~/.dsh/speech-to-text/gigaam-ru/models/gigaam-v3/model.onnx         319 869 121 B
~/.dsh/speech-to-text/gigaam-ru/models/gigaam-v3/tokens.txt               2 006 B
~/.dsh/speech-to-text/gigaam-ru/models/silero/silero_vad.onnx         1 807 522 B

Use Download and prepare in the voice settings, or just press the microphone: the provider wakes the worker and loads the model (≈0.4 s) once the cache is complete.

Offline install

Copy the cache directory from a prepared machine to the same path on the target machine (the provider re-verifies every file against the pinned SHA-256 before use), or point the provider at existing files through modelDirectory / vadModelPath in your profile patch. To forbid public downloads entirely, set modelOrigin to an internal Hugging Face-compatible mirror (see Configuration).

Use

1. Press the microphone in the message input, dictate in Russian, press Stop.
2. The transcript is inserted into the draft — with punctuation, capitalization and normalized numbers (в 15:30, 9:00), ready to edit or send.
3. Switch recognizers/languages any time in Settings → Voice input; the selection is stored per profile by DSH's Settings service.

Quality (measured on the reference set)

	GigaAM-v3 e2e CTC INT8 (default)	Whisper Small INT8 (fallback)
Corpus WER	3.03 % (0 lexical errors)	13.64 %
Exact normalized matches	4 / 5 clips	2 / 5
RTF @ 2 threads	0.017	0.257
RTF @ 4 threads	0.014	0.264
Peak RSS	564 MiB	1.40 GiB
Load	0.36 s	0.61 s
Punctuation / ITN	yes / yes	yes / yes

Reference clips, per-clip transcripts and the full benchmark are in docs/asr-validation.md and docs/model-assets.json.

Configuration

The provider row in a profile is speech-to-text-gigaam. Override it in ~/.dsh/profiles/<profile>/cordis.patch.yml (the user patch always wins over the bundle layer):

- id: speech-to-text-gigaam
  config:
    threads: 4              # CPU threads (default 2)
    precision: int8         # int8 (default) or fp32 (larger, not pinned yet)
    engine: gigaam-ctc      # gigaam-ctc (default) or whisper
    idleTimeoutMs: 300000   # unload the worker after 5 min of silence
    # modelDirectory: /abs/path/to/models/gigaam-v3
    # vadModelPath: /abs/path/to/silero_vad.onnx
    # modelOrigin: https://hf-mirror.com     # single explicit mirror, no fallback

Key	Default	Meaning
providerId	gigaam-ru	Id the provider registers under (shown in the selector)
engine	gigaam-ctc	gigaam-ctc (Russian) or whisper (multilingual fallback)
precision	int8	int8 / fp32 (only INT8 assets are pinned today)
threads	2	Native inference threads (4 gives ~20 % on this CPU)
dataRoot	~/.dsh/speech-to-text/gigaam-ru	Absolute model cache root
modelDirectory, vadModelPath	–	Use existing absolute files instead of downloading
modelOrigin, modelOrigins	Hugging Face + HF-Mirror	Download sources (single modelOrigin disables fallback)
segmentSeconds	30	Max VAD segment duration
vadThreshold, minSpeechSeconds, minSilenceSeconds, vadOnsetPaddingSeconds	0.5, 0.25, 0.5, 0.4	Silero VAD segmentation tuning
idleTimeoutMs	300000	Stop the worker after inactivity (0 keeps it)
maxPending, inferenceTimeoutMs, prepareTimeoutMs	4, 120000, 3600000	Queue and time bounds

The service-level default selection lives on the speech-to-text row (defaultProvider: gigaam-ru, language: ru).

Engines

• gigaam-ctc (default) — GigaAM-v3 e2e CTC converted for sherpa-onnx NeMo-CTC. Russian, MIT-licensed, best quality and speed. English is not advertised: GigaAM transliterates English words into Cyrillic and makes spelling errors.
• whisper — Whisper Small INT8 (multilingual, 99 languages). Slower and less accurate on Russian, but useful for mixed Russian/English dictation. Set engine: whisper and choose language ru, en or auto in the voice settings. (Note: sherpa-onnx accepts only "" for auto-detect; the provider maps auto to "" so the native layer never sees the literal string.)

How it works

Host plugin (lib/index.js)                     worker subprocess (lib/worker.js)
─────────────────────────                      ──────────────────────────────────
inspect cache on activation (no downloads)     load sherpa-onnx OfflineRecognizer
prepare(): download → verify → load            load Silero VAD
transcribe(): POST WAV to 127.0.0.1:PORT  ───► decode VAD segments (GigaAM or Whisper)
  bearer token, one serial worker               → JSON transcript back
idle timeout / cancellation → terminate

• Assets are revision-pinned on Hugging Face with size + SHA-256 verification while streaming; a verified file is never downloaded again.
• The native sherpa-onnx-node runtime comes from the package dependency (or the copy inside the DSH app bundle as a fallback) — this package never bundles platform binaries.
• The worker is a Node child process with a random 32-byte bearer token; the loopback listener only accepts POST /transcribe from the Host. Audio stays in memory.
• The provider is a normal DSH plugin entry (apply/inject) with a preparation state machine (check → download → verify → load), so the standard voice UI cards, progress, cancellation and settings work unchanged.

Troubleshooting

Symptom	Fix
No microphone in the composer	Enable Plugins → Voice input; refresh the window (Cmd+R).
Provider not in the selector	Check Plugins for a load error on this bundle; make sure Voice input is enabled (the provider needs the speechToText service) and the profile lists the bundle after it.
"Download and prepare" errors	Check network/disk; the UI shows source and reason. modelOrigin can point at a mirror.
Cache from another machine ignored	File names must match runtime/assets.json; the provider verifies filename, size and SHA-256.
Slow / high CPU	Lower segmentSeconds, keep threads: 2; the worker unloads after idleTimeoutMs.
Need to remove	dsh plugin --profile desktop remove dsh-speech-to-text-ru and toggle Voice input off (or leave it for SenseVoice).

Privacy

Model weights are downloaded from Hugging Face; nothing else touches the network. Recordings are transcribed on the Host CPU, in memory, by a local process, and are never uploaded. No API keys, no telemetry from this plugin.

Development

npm install                      # only the pinned sherpa native runtime
node --test test/smoke.test.mjs  # 15 model-free contract/unit tests (CI runs these)
npm pack                         # produces the installable tarball
node scripts/transcribe.mjs --wav samples/16k.wav --model-dir ~/.dsh/speech-to-text/gigaam-ru/models

Dev install from a checkout (file:// module URL, no package install):

dsh plugin --profile dev-profile add /path/to/checkout

Upstream / official bundle

DeepSeek Harness' CONTRIBUTING.md currently states that external pull requests are not accepted, so instead of opening a PR the integration proposal was posted in the project's discussions:

• Upstream proposal (Ideas): https://github.com/deepseek-ai/deepseek-harness/discussions/11745
• Plugin announcement (Show Your Plugins!): https://github.com/deepseek-ai/deepseek-harness/discussions/11746

The implementation-ready change set — new packages/experimental/speech-to-text-gigaam package mirroring speech-to-text-sensevoice, the voice-input-bundle patch row, official-bundle manifests, docs and tests — is described in docs/upstream.md. If the project opens external contributions (or the maintainers ask for the code), the port is mechanical and this repository is the working prototype.

Credits & licenses

• GigaAM-v3 — © SaluteDevices / Sber, MIT. ONNX conversion for sherpa-onnx by Smirnov75.
• sherpa-onnx — Apache-2.0 (next-gen Kaldi team); Silero VAD is MIT.
• DeepSeek Harness — the provider architecture (preparation state machine, worker transport, VAD segmentation, plugin contract) is adapted from the MIT-licensed @deepseek-ai/dsh-experimental-speech-to-text-sensevoice package and the documented speechToText service contract.
• This bundle: MIT, see LICENSE.

---

Русская документация: README.ru.md