dsh-speech-to-text-gigaam

Local Russian speech recognition for DeepSeek Harness voice input, powered by Sber's GigaAM v3 E2E RNNT and running entirely on the Host CPU. Punctuation and capitalisation are part of the output.

DSH ships a single recognizer — sensevoice-local (SenseVoiceSmall) — whose language hints are auto, zh, en, yue, ja, ko: no Russian. This bundle registers an additional speechToText provider, gigaam-local, and selects it as the default.

Audio never leaves the machine. The only network access is the one-time download of the model weights from Sber's public CDN.

Features

• Russian with punctuation. GigaAM v3 E2E RNNT returns punctuated, capitalised text — no post-processing heuristics.
• Guided download in the UI. The provider never downloads on activation: the voice settings show an explicit “Download and prepare” card with disk, memory and time estimates, a progress bar, cancellation and checksum verification.
• One model, no choices. v3_e2e_rnnt is the only checkpoint the provider installs. Other GigaAM models can be pre-downloaded with setup.sh and selected in the row config, but the UI will not offer them.
• Fully local. No API keys, no cloud transcription; inference runs in a Python worker on the Host.
• Long recordings. GigaAM accepts at most 25 s per call; the worker cuts longer audio at low-energy points and joins the chunk transcripts.
• Cancellable and serialized. Cancel discards the recording and kills the worker; parallel recordings queue instead of loading several models at once.
• Portable defaults. Paths resolve to ~/.dsh/speech-to-text/gigaam; no absolute paths are baked into the bundle.

Requirements

• DeepSeek Harness with the voice-input bundle (@deepseek-ai/dsh-experimental-voice-input-bundle).
• ffmpeg on PATH (brew install ffmpeg).
• Python 3.10+ with torch, torchaudio and gigaam — python/setup.sh builds it.
• About 430 MB for the model and 2.5 GB for the Python environment.

Install

1. Prepare the Python environment (once). It creates a venv in ~/.dsh/speech-to-text/gigaam, installs PyTorch and GigaAM, then downloads and warms up the model:

   ./python/setup.sh
   

   Override the location with GIGAAM_DATA_ROOT=/path ./python/setup.sh. The model is optional here: if the script is run without network access, the UI can download it later.

2. Install the bundle. In the Harness UI: Plugins → Add plugin, paste the repository URL (https://github.com/kelvmg/dsh-speech-to-text-gigaam), this directory's absolute path, or the published package name — then Install and Enable now.

3. Pick the recognizer. Open the voice-input bundle details (or Settings → voice input), choose GigaAM v3 · Russian (punctuation) and language auto or ru.

Guided model download

On activation the provider only inspects the cache. With no model on disk it reports unprepared, and the voice settings show a card with:

Estimate	Value
Disk	~0.5 GB
Memory	~1.5 GB
Time	1–5 minutes

“Download and prepare” then runs four steps:

1. Check local resources — the Python interpreter and the worker script exist.
2. Prepare recognition model — downloads v3_e2e_rnnt.ckpt (428 MB) and v3_e2e_rnnt_tokenizer.model (0.25 MB) from https://cdn.chatwm.opensmodel.sberdevices.ru into downloadRoot.
3. Verify model files — size, MD5 of the checkpoint (2730de7545ac43ad256485a462b0a27a) and SHA-256 of the tokenizer.
4. Load GigaAM v3 · Russian (punctuation) — a warm-up run, so the first dictation takes seconds instead of half a minute.

The download can be cancelled at any point; a partial file is removed and the next attempt starts over. On success the provider reports standby/ready and the microphone records normally.

If the interpreter is missing, the card reports a failure with the instruction to run python/setup.sh — nothing is downloaded or recorded in that state.

Configuration

The bundle patch declares the row speech-to-text-gigaam:

Key	Default	Meaning
dataRoot	~/.dsh/speech-to-text/gigaam	Base directory for venv/ and models/
pythonPath	<dataRoot>/venv/bin/python	Interpreter with gigaam installed
scriptPath	<package>/python/gigaam_stt.py	Worker script
downloadRoot	<dataRoot>/models	Model cache
model	v3_e2e_rnnt	Only this model is offered for guided download
device	cpu	cpu, mps or cuda; CPU is faster than MPS on Apple silicon
maxChunkSeconds	20	Chunk length for recordings over 25 s (clamped to 3–24.5)
timeoutMs	900000	Worker deadline, including the first model download
providerId	gigaam-local	Registered provider id
providerName	GigaAM v3 · Russian (punctuation)	Label in the provider list
downloadSources	CDN Sber	Origins offered for the download
loadModel	true	Skip the warm-up step (tests)

How it works

composer mic → speechToText service → gigaam-local provider (lib/index.js)
             → preparation (lib/preparation.js) for the model files
             → Python worker (python/gigaam_stt.py) → GigaAM on CPU → text

The worker is spawned per recording, so the provider is stateless; the model loads in about 2 s and inference is roughly 0.04× real time on an M3.

Performance (Apple M3, 8 cores, CPU)

Recording	Wall time
10.7 s	~3 s
39.5 s (2 chunks)	~3 s
Warm-up (verify + load)	~3 s

Limitations

• One model. Only v3_e2e_rnnt is downloadable from the UI. Other GigaAM models (multilingual, CTC, SSL) must be fetched manually and referenced in the config.
• No streaming. The DSH speech service transcribes complete recordings only, so text appears after Stop.
• Python environment is scripted. prepare() downloads weights, but the 2.5 GB Python environment still comes from python/setup.sh.
• CPU memory. Expect roughly 1.5 GB RSS while a transcription runs.

Troubleshooting

• ffmpeg is required but was not found — install ffmpeg; the worker prepends /opt/homebrew/bin, /usr/local/bin and /opt/local/bin to PATH already.
• Python environment not found … Run python/setup.sh first — the venv is missing or pythonPath points elsewhere.
• Downloaded … failed its checksum; try again — a proxy or mirror returned corrupted bytes; the partial file is removed, retry the download.
• The worker writes diagnostic lines and full tracebacks to <tmpdir>/dsh-gigaam-stt.log, because the desktop app has no stderr console.

Privacy

Recordings are written to a temporary directory, transcribed locally and deleted immediately. No audio, transcript or telemetry is sent anywhere. Setup downloads PyTorch wheels from your pip index and GigaAM weights from cdn.chatwm.opensmodel.sberdevices.ru.

License

MIT © 2026 Evgeny Klimov. GigaAM itself is MIT-licensed by SberDevices; see salute-developers/GigaAM.

Changelog

• 0.3.0 — guided model download with progress, cancellation and checksum verification; explicit setup card; single-model scope (v3_e2e_rnnt); preparation unit tests.
• 0.2.0 — returns the { text, audioSeconds, inferenceSeconds } result shape the voice API requires; portable defaults instead of absolute paths; PATH fix for GUI-launched apps; chunk-length clamping; smoke tests; package metadata.
• 0.1.0 — first working provider.