Садо (Sado) - Tajik offline voice-to-text for Windows
How to get access: please fill in the short form at saidzoda.com/access. We approve requests after we receive the form.
Install (non-technical, one click)
- Download Sado-Setup.exe.
- Double-click it -> Install -> 'Садо' appears in the Start menu + desktop.
- Open Садо, click the mic, speak Tajik -> text. Fully offline, on-device (nothing is uploaded).
Speech->text with real trained punctuation (PERIOD-F1 0.984) + number/currency, ~11-16% WER. Unsigned build for now: Windows SmartScreen -> More info -> Run anyway (a signed build removes this).
(Advanced: TajikSTT-windows.zip is the unpacked app folder — unzip and run Sado.exe.)
Bundled model (standalone copy)
model/whisper-ca-ep4-ct2/ — the CTranslate2 (int8) Whisper-CA epoch-4 build that the
Saidzoda Lab Sado voice stack runs in production for Tajik speech-to-text. 1.5 GB, runs fully
offline on CPU or GPU via ctranslate2 / faster-whisper.
Added 2026-09-24 so this repository contains both the application and the model as a standalone copy (previously the model existed only baked inside the app bundles).
from faster_whisper import WhisperModel
m = WhisperModel("model/whisper-ca-ep4-ct2", device="cpu", compute_type="int8")
segments, info = m.transcribe("audio.wav", language="tg")
On accuracy: the Whisper-CA family's audited Tajik benchmark figure is ~9.8 % WER (ckpt-86544, full FLEURS-tg test set). A 3-clip load-time smoke check that ships with the serving directory shows 6.78 % — that is an n=3 sanity probe, not a benchmark, and should not be quoted as one. The app-level figure quoted above (~11-16 % WER) is end-to-end, including punctuation and number formatting.