dsh-gigaam-stt

Local Russian speech-to-text for the DeepSeek Harness chat. Dictate Russian, get Cyrillic text — with capitalisation and punctuation — transcribed on your CPU, offline, with no API keys and no cloud calls.

The engine is GigaAM v3 (220M Conformer, MIT) exported to ONNX (e2e_ctc, int8) and run by sherpa-onnx. The plugin plugs into DSH's existing voice-input pipeline: it registers a speechToText provider, so the microphone button in the chat composer starts working with Russian.

mic → 16 kHz mono PCM16 WAV → this provider → worker thread → sherpa-onnx → "Привет, как дела?"

Features

• Russian only, properly. Per ai-sage's benchmark, GigaAM v3 reports ~8.4 % WER on Russian against ~19 % for T-One and ~25 % for Whisper-large-v3 — and its e2e variants emit capitalisation and punctuation directly.
• CPU-first and small. 320 MB int8 model, ~420 MB RSS while loaded, ~0.04 RTF — 11 s of speech is transcribed in about 0.45 s on four threads.
• Lazy and self-cleaning. The model downloads on first use, loads on first recording (~1 s), unloads after five idle minutes. Nothing runs in the background.
• No dependencies. The plugin itself imports only Node built-ins; sherpa-onnx is reused from your DSH installation. No build step, no native compilation, no Python.
• Long recordings work. Audio is cut into pause-aligned windows of 20 s before decoding: a single pass over anything past ~200 s aborts sherpa-onnx (5 000 encoder frames), while pause-aligned windows keep memory at ~520 MB and stay within the model's trained length. 214 s of audio decodes in 8.7 s.
• Numbers as digits. "двадцать пять" → "25", "в две тысячи двадцать пятом году" → "в 2025 году". Conservative: single words below two stay as words, so "ни один" and "один в один" survive. Switch off with numbers: words.
• Non-blocking. Recognition runs in a worker_threads worker, so the chat never stalls while decoding.
• Safe downloads. Pinned URL, size and SHA-256 for every asset; interrupted downloads resume from .part files.

Requirements

• DeepSeek Harness with the voice-input stack — @deepseek-ai/dsh-experimental-voice-input-bundle (speech service, mic UI and Remote API). It ships with DSH's experimental voice input; without it this bundle installs but has no microphone to attach to.
• Node.js 22+ (DSH itself needs a modern Node), x86-64 or arm64 Linux/macOS/Windows.
• ~350 MB free disk for the model, ~450 MB RAM while it is loaded.
• A browser with microphone access to DSH Web (the mic permission prompt appears on first recording).

Install

Ask your DSH agent to install the bundle, or use the Plugin Manager panel in the DSH Web UI — the install target accepts any spec pnpm accepts:

Source	Install target
npm registry	dsh-gigaam-stt
GitHub repository	github:OldSchool686/dsh-gigaam-stt
Local checkout	absolute path, e.g. /home/me/dsh-gigaam-stt
Tarball	/home/me/dsh-gigaam-stt-1.1.0.tgz

The bundle's patch does three things: it makes gigaam-local the default speech provider with language ru, disables the bundled SenseVoice provider (its languages are zh/en/yue/ja/ko, so it would only add noise), and mounts this provider with dataRoot: !!js dshHomePath('speech-to-text', 'gigaam').

If your profile has no voice-input bundle, the two row overrides are reported as warnings and skipped — install @deepseek-ai/dsh-experimental-voice-input-bundle first, then re-apply this bundle.

First run

Press the microphone button in the chat composer. The preparation card shows a one-time download of two files (~320 MB) from Hugging Face, verified by SHA-256, and the state changes downloading → loading → ready.

Assets are stored in ~/.dsh/speech-to-text/gigaam/:

File	Bytes	SHA-256
gigaam_v3_e2e_ctc_int8.onnx	319869121	0aacb41f70f0f5aaac4b45dd430337b9e16b180f22c72af04db8516e7609c3c0
gigaam_v3_e2e_ctc_tokens.txt	2006	f8eb9b115e2748db9c40a5897cae11dd0678cc0b40fd7e25f8c43b3bf28715e4

To pre-seed them (air-gapped machine, CI image, shared cache), put the two files there with those names and sizes — an already-present, correctly-sized file is used as is:

mkdir -p ~/.dsh/speech-to-text/gigaam
curl -L -o ~/.dsh/speech-to-text/gigaam/gigaam_v3_e2e_ctc_int8.onnx \
  https://huggingface.co/Smirnov75/GigaAM-v3-sherpa-onnx/resolve/main/gigaam_v3_e2e_ctc_int8.onnx
curl -L -o ~/.dsh/speech-to-text/gigaam/gigaam_v3_e2e_ctc_tokens.txt \
  https://huggingface.co/Smirnov75/GigaAM-v3-sherpa-onnx/resolve/main/gigaam_v3_e2e_ctc_tokens.txt

modelOrigins defaults to https://huggingface.co first and https://hf-mirror.com second; the UI shows an origin picker when both are reachable.

Settings

Configure the speech-to-text-gigaam row in your profile patch (or via the plugin's settings UI):

Option	Default	Meaning
providerId	gigaam-local	id under which the provider registers; speech-to-text.defaultProvider must match it.
dataRoot	required	absolute directory for the model files.
modelOrigins	["https://huggingface.co", "https://hf-mirror.com"]	download mirrors, probed in order.
threads	4	sherpa-onnx intra-op threads. Lower it on small CPUs, raise it for throughput.
idleTimeoutMs	300000	unload the model after this much inactivity.
maxAudioSeconds	600	reject longer recordings before decoding.
maxSegmentSeconds	20	decode in pause-aligned windows of at most this length (5–180). Longer audio is split, not rejected.
numbers	digits	digits rewrites Russian number words to digits; words leaves them as spoken.
progressIntervalMs	100	download progress publication interval.

Example override:

- id: speech-to-text-gigaam
  config:
    dataRoot: !!js dshHomePath('speech-to-text', 'gigaam')
    threads: 2

How it works

1. The DSH speech service validates that the recording is a canonical 16 kHz mono PCM16 WAV and hands the bytes to the provider.
2. The provider serialises requests, wakes the worker if needed, and transfers the audio buffer.
3. The worker strips the 44-byte WAV header, converts PCM16 to float32, and splits the speech into windows of at most maxSegmentSeconds, cutting inside pauses (adaptive RMS threshold over 10 ms frames) and dropping leading and trailing silence. Each window goes through OfflineRecognizer with modelConfig.nemoCtc (GigaAM e2e_ctc emits punctuation and casing directly).
4. Windows are joined, number words are normalised (unless numbers: words), and the text comes back with audioSeconds and inferenceSeconds; the model unloads after the idle timeout.

Measured on an 8-vCPU box (Node 24, int8, 4 threads): model load 0.84 s, decode 0.44–0.46 s for 11.29 s of audio, RTF ~0.043 for long input (214 s in 8.7 s), RSS ~420 MB with the model resident and ~520 MB during a long transcription. One pass over 214 s without windowing used ~2 GB and aborted the process.

Self-test without the UI

cd dsh-gigaam-stt
node scripts/selftest.mjs                                     # 1 s of silence must decode to ""
GIGAAM_SAMPLE=/path/to/speech.wav node scripts/selftest.mjs   # expect Cyrillic

The script talks to worker.js directly; it resolves sherpa from your DSH install (override with GIGAAM_SHERPA_PATH) and reads the model from ~/.dsh/speech-to-text/gigaam.

Troubleshooting

Symptom	Cause and fix
sherpa-onnx-node не найден	The plugin could not find the addon. It searches node_modules/sherpa-onnx-node upward from argv[1], process.execPath, the plugin directory and cwd. Point it explicitly: GIGAAM_SHERPA_PATH=/abs/path/sherpa-onnx-node/sherpa-onnx.js in the DSH process environment.
No microphone button / no provider in the list	The voice-input bundle is missing or disabled; install @deepseek-ai/dsh-experimental-voice-input-bundle.
Download fails behind a firewall	The second origin https://hf-mirror.com is tried automatically; otherwise pre-seed dataRoot manually (see above).
Audio must be a canonical 16 kHz mono PCM16 WAV recording	The client sent another format; this provider accepts only that one.
First recording shows “Разбужаем локальный голос…”	The model was unloaded after idleTimeoutMs; loading takes ~1 s.
Numbers stay as words (or an odd rewrite)	numbers: digits is the default; set numbers: words to keep them spoken, or raise maxSegmentSeconds if a phrase was split mid-number.
A number came out wrong, or you want to know what the recognizer heard	Set GIGAAM_DEBUG=1 in the DSH process environment: the worker then logs every window's raw text and the normalised result ([gigaam-stt] debug: ...) to the harness log, which separates a mishearing from a normalisation bug.
Recording is cancelled when you switch to another window	The harness voice-input client cancels an in-flight capture on window blur and on a hidden tab (client.js, blur/visibilitychange handlers). Keep the chat window focused while dictating; this provider never sees the discarded audio.
Recording stops by itself after ~2 minutes	The transport caps recordings at 120 s / 4 MiB by default; this bundle raises it to 600 s / 20 MiB. For longer single recordings raise maxDurationSeconds + maxAudioBytes on api-speech-to-text and maxAudioSeconds here.
Preparation stuck in failed	Fix the reported cause and record again — the next recording retries the worker automatically.

Uninstall

Remove the bundle through the Plugin Manager (or ask your agent). The model stays in ~/.dsh/speech-to-text/gigaam — delete that directory to reclaim ~320 MB.

Privacy

• No credentials. The plugin reads one environment variable, GIGAAM_SHERPA_PATH (a local filesystem path). It never touches DSH credentials, API keys, cookies or the settings store.
• No telemetry, no analytics, no update checks, no accounts.
• The only network access is the one-time model download over HTTPS from modelOrigins: a plain GET with no authorization header. It can be skipped entirely by pre-seeding dataRoot.
• Audio never leaves the machine. The recording goes to a local worker thread, is decoded on the CPU, and is dropped afterwards.
• Nothing is written outside dataRoot (plus .part files inside it while a download is in progress).

License and credits

• This bundle: MIT (see LICENSE).
• GigaAM v3 by ai-sage — MIT; the ONNX conversion used here comes from the community repo Smirnov75/GigaAM-v3-sherpa-onnx.
• sherpa-onnx by k2-fsa — Apache-2.0.
• The voice-input UI and the speechToText service are part of DeepSeek Harness.

---

Кратко по-русски

Плагин добавляет в чат DeepSeek Harness локальное распознавание русской речи: жмёте микрофон, говорите, в поле ввода появляется текст с пунктуацией и заглавными буквами. Всё считается на процессоре, без интернета после первой загрузки модели (320 МБ), память ~450 МБ пока модель загружена, выгрузка через 5 минут простоя.

Установка: попросите агента установить бандл dsh-gigaam-stt (или github:OldSchool686/dsh-gigaam-stt), либо укажите абсолютный путь к этому каталогу в Plugin Manager. Нужен установленный голосовой ввод DSH (@deepseek-ai/dsh-experimental-voice-input-bundle). Первый клик по микрофону скачивает модель в ~/.dsh/speech-to-text/gigaam. Настройки: threads, idleTimeoutMs, modelOrigins, dataRoot.