ScribeAir — Offline Speech-to-Text for Windowsv1.3.0

Voice Input Without Cloud or Subscriptions

ScribeAir 1.1.0 is a free speech recognition tool for Windows that works entirely offline. Hold a hotkey, speak — and the text instantly appears in any application: messenger, editor, IDE, or browser.

No data ever leaves your computer. No accounts, no subscriptions, no time limits.

---

What's New in 1.1.0

• GigaAM — new ASR engine for Russian language. 3.3% WER (word error rate) vs 32.6% with Whisper base — a 90% error reduction.
• Three recognition engines to choose from: GigaAM (quality), Whisper (multilingual), Vosk (speed).
• 93 automated tests — full coverage of all components: transcription, correction, filtering, pipeline.
• ASR backend switching from tray — change engine and model without restarting.
• Audio device selection — switch microphone from the tray menu.

---

Testing Results

Recognition quality was benchmarked on an audiobook corpus (Russian literary text of varying complexity): short phrases 6–7 seconds, long passages 25–30 seconds, rare words.

Recognition Quality (WER — Word Error Rate)

Engine	WER	Error Reduction	Speed	Notes
GigaAM v3-e2e-rnnt	3.3%	90.0%	0.66s	Best quality + speed, punctuation
GigaAM v3-rnnt	3.3%	90.0%	0.82s	Best quality
GigaAM v3-e2e-ctc	4.2%	87.2%	1.08s	Good balance
Whisper large-v3-turbo (GPU)	7.9%	75.7%	0.44s	Multilingual
Vosk small	13.0%	60.0%	0.75s	Minimal size (50 MB)
Whisper base (CPU)	32.6%	—	0.45s	v1.0 baseline

GigaAM on CPU outperforms all Whisper models on GPU in Russian text quality (3.3% vs 7.9% WER).

Automated Testing

Category	Tests	Coverage
Transcription (Whisper)	15	RU/EN, streaming, multilingual, hints, silence
Full Pipeline	9	Audio → ASR → correction → text, RU→EN translation
T5 Correction	11	Error correction, chain-of-correction, model unloading
Hallucination Filtering	14	Known phrases, delooping, duration anomalies
Audio Processing	11	Normalization, silence trimming, high-pass filter, edge cases
Streaming Pipeline	7	VAD, lifecycle, chunk processing
Integration	8	Silero VAD, noisereduce, config v2
Total	93	100% passing

---

Key Advantages

• 100% Offline & Private — voice is processed locally, never sent to the cloud. Complete privacy out of the box.
• 90% Error Reduction — GigaAM (Sber, 700K hours of training data) delivers 3.3% WER for Russian language.
• Three ASR Engines — GigaAM for quality, Whisper for multilingual, Vosk for speed.
• GPU & CPU — CUDA acceleration for maximum speed, CPU build for any PC.
• 93 Automated Tests — every component is verified: from audio processing to the full pipeline.
• Free & Open Source — MIT license. Use, modify, and distribute without restrictions.

---

How It Works

1. Launch ScribeAir — a microphone icon appears in the system tray
2. Hold the hotkey (default: LShift + RShift)
3. Speak — text appears on screen in real time
4. Release the key — finished text is inserted into the active field

Models are downloaded automatically on first launch. After that, no internet is needed.

---

Recognition Modes

Mode	Engine	Languages	Quality (WER)	Speed
Auto (default)	GigaAM (CPU) / Whisper (GPU)	RU / Multi	3.3% / 7.9%	0.66s / 0.44s
GigaAM	GigaAM ONNX	Russian only	3.3%	0.66s
Whisper	faster-whisper	RU, EN, auto, translation	7.9–32.6%	0.44–2.3s
Vosk	Vosk offline	Russian	13.0%	0.75s

---

Features

Speech Recognition

• GigaAM (Sber) — ONNX model trained on 700K hours of Russian speech. 6 model variants.
• faster-whisper (CTranslate2) — optimized Whisper. Models: base, small, medium, large-v3-turbo.
• Vosk — compact offline engine (50 MB), ideal for short phrases.
• Streaming pipeline with Silero VAD — text appears as you speak
• Whisper hallucination filtering (known artifacts are removed automatically)

Text Processing

• T5 correction (bond005/ruT5-ASR-large) — fixes ASR errors for Russian language
• Chain-of-Correction for complex sentences
• Custom vocabulary for professional terminology

Audio & Interface

• Audio preprocessing: pre-emphasis, normalization, 80 Hz high-pass filter, noise reduction, silence trimming
• Configurable hotkeys (LShift+RShift, Win+Shift, etc.)
• Semi-transparent overlay window with transcription progress
• Microphone and ASR backend selection from tray menu
• Windows autostart

---

Two Build Variants

Variant	Size	Description
CUDA	~2.4 GB	Full GPU support. Maximum speed on NVIDIA graphics cards.
CPU	~800 MB	Works on any PC without a graphics card.

AI models are downloaded automatically on first launch and cached locally.

---

Comparison with Alternatives

Feature	ScribeAir 1.1	CamoVoice	Handy	Speechnotes
Price	Free	$24.99	Free	Freemium
Offline	Yes	Yes	Yes	No
WER (Russian)	3.3%	N/A	N/A	N/A
ASR Engines	3 (GigaAM, Whisper, Vosk)	1	1	1 (cloud)
Open Source	MIT	No	Yes	No
GPU Acceleration	CUDA	No	No	Cloud
Text Correction	T5 Neural Net	No	No	No
Automated Tests	93 tests	N/A	N/A	N/A
Mixed RU+EN	Yes	No	No	No
Streaming	Real-time	No	No	Yes

---

System Requirements

	Minimum	Recommended
OS	Windows 10/11	Windows 11
RAM	8 GB	16 GB
GPU	Not required (CPU build)	NVIDIA 6+ GB VRAM
CUDA	—	CUDA 12.x
Disk	~3 GB	~5 GB (with models)

---

Technology Stack

• GigaAM (Sber) — ONNX model for Russian ASR, trained on 700K hours of speech
• faster-whisper — optimized Whisper engine based on CTranslate2
• Vosk — compact offline ASR engine (Kaldi)
• Silero VAD — neural voice activity detector
• T5 (bond005/ruT5-ASR-large) — correction model for Russian ASR
• ONNX Runtime — cross-platform inference for GigaAM
• PyTorch + CUDA — GPU-accelerated inference
• PyInstaller — standalone EXE packaging