Windows Offline Transcribe

WinUI 3 (Windows App SDK) desktop app for offline speech transcription. All ASR inference runs locally after model download.

This repo also supports an optional Translate mode (offline MT + optional TTS) in the same app.

Screenshots

File Transcription Demo

Select an audio file (.wav / .mp3) and get the transcript in seconds — all processing runs locally.

Current Scope (Code-Accurate)

• Live transcription with confirmed text plus rolling hypothesis.
• Two UI modes (left-nav):
  • Transcribe: transcription only.
  • Translate: transcription + offline translation + optional TTS playback.
• Audio source switching: Microphone, System Audio (WASAPI loopback).
• In-app model download/load/switch (12 models across 5 engine types).
• File transcription (.wav, .mp3).
• Runtime stats while recording (CPU, RAM, tok/s, elapsed audio). Note: tok/s is a rough word-per-second estimate.
• History: saves transcript + session audio; export a shareable ZIP (text + metadata + audio).
  • Translate mode records may also include translation.txt and tts.wav in the exported ZIP.
• Diagnostics: Evidence Mode (events.jsonl + model manifests + screenshots; export one ZIP).

Supported Models & Benchmarks

Defined in src/OfflineTranscription/Models/ModelInfo.cs. Models are downloaded from Hugging Face at runtime and stored under %LOCALAPPDATA%\\OfflineTranscription\\Models\\<model-id>\\.

Benchmark: JFK inauguration excerpt (11s, 22 words). Device: i5-1035G1 (4C/8T), 8 GB RAM, CPU-only.

• Params (weights) = number of trainable model parameters (e.g. 600 M = 600 million parameters).
• Disk (download) = approximate total download size for all model files (ONNX weights, tokenizer, config).
• RTF = Real-Time Factor (lower = faster). RTF < 1.0 means faster than real-time.
• Bold = recommended models.

Model weights are not distributed with this repo; model licensing varies. See NOTICE.

Want to see a new model or device benchmark? If there is an offline ASR model you would like added, or you have benchmark results to share from your hardware, please open an issue. Community contributions of inference benchmarks on different Windows devices are welcome.

Engines / Inference Methods

This app has multiple inference paths:

Live Recording (Microphone / Loopback)

• sherpa-onnx streaming (Zipformer):
  • True streaming decode (100 ms audio chunks).
  • Endpoint detection commits "confirmed" text on silence and resets the stream.
• whisper.cpp and sherpa-onnx offline:
  • Chunk-based loop while recording.
  • Window size: ~15 s for whisper.cpp, ~3.5 s for sherpa-onnx offline.
  • Optional VAD: skips work when recent audio energy is below a threshold.
  • Prefix-match between consecutive results to split confirmed text vs rolling hypothesis.

Provider selection:

• sherpa-onnx offline probes DirectML first and falls back to CPU.
• sherpa-onnx streaming uses CPU for stable real-time behavior.

File Transcription

• Uses NAudio to decode audio, resamples to 16 kHz mono Float32, and runs a single engine pass.

Qwen-ASR (C engine)

• qwen-asr (antirez/qwen-asr): native C engine for Qwen3-ASR models.
• Uses memory-mapped BF16 safetensors with OpenBLAS for inference.
• Requires compiling qwen_asr.dll from source (see Native Dependencies).

Windows Speech API

• WindowsSpeech: built-in Windows recognizer via WinRT SpeechRecognizer.
• Works offline if the Windows language pack is installed.
• Limited accuracy compared to neural ASR models. No native DLLs required.

Translate Mode

When you select Translate in the left navigation:

• The app runs offline machine translation (CTranslate2) on the live transcript.
• Translation is always active in Translate mode (there is no separate enable toggle).
• Optional: enable Speak Translated Audio (Windows.Media TTS). The app will stop recording during TTS playback to avoid feedback loops, then resume.

Translation settings:

• Configure Source Language and Target Language in Settings (Translate mode only).
• Translation models are cached under %LOCALAPPDATA%\\OfflineTranscription\\TranslationModels\\.

Legacy Migration (Windows)

On first run, the app performs a best-effort migration from the legacy Windows app data folder %LOCALAPPDATA%\\OfflineSpeechTranslation\\ into %LOCALAPPDATA%\\OfflineTranscription\\ when the new location does not already exist.

Prereqs

• Windows 10/11 (x64)
• Visual Studio 2022 or newer
• .NET 8 SDK

Native Dependencies

This app uses native engines via P/Invoke:

• whisper.cpp: whisper.dll (see src/OfflineTranscription/Interop/WhisperNative.cs)
• sherpa-onnx C API: sherpa-onnx-c-api.dll (see src/OfflineTranscription/Interop/SherpaOnnxNative.cs)
• qwen-asr: qwen_asr.dll + libopenblas.dll (see src/OfflineTranscription/Interop/QwenAsrNative.cs)
• offline translation (CTranslate2 wrapper): OfflineTranscription.NativeTranslation.dll (see src/OfflineTranscription/Interop/NativeTranslation.cs)
• Windows Speech: no native DLLs needed (WinRT API)

Place required .dll files in:

• libs/runtimes/win-x64/

They are copied to the build output directory by src/OfflineTranscription/OfflineTranscription.csproj.

If you use the sherpa-onnx release bundle, make sure you also include its dependent DLLs (for example onnxruntime.dll and DirectML-related DLLs if applicable).

Building qwen_asr.dll (Windows)

For qwen-asr, compile qwen_asr.dll from antirez/qwen-asr source and place alongside libopenblas.dll (from OpenMathLib/OpenBLAS releases).

Building OfflineTranscription.NativeTranslation.dll (Windows)

The native translation wrapper lives in src/OfflineTranscription.NativeTranslation/ and is built with CMake.

1. Install a C++ toolchain (Visual Studio Build Tools) and CMake.
2. Build the DLL using the instructions in src/OfflineTranscription.NativeTranslation/README.md.
3. Copy the produced OfflineTranscription.NativeTranslation.dll (and any dependency DLLs) into:
   • libs/runtimes/win-x64/

Build

Open OfflineTranscription.sln and build x64 (Debug or Release).

Diagnostics

• Save Screenshot writes PNGs under %LOCALAPPDATA%\\OfflineTranscription\\Diagnostics\\
• Evidence Mode writes under %LOCALAPPDATA%\\OfflineTranscription\\Evidence\\ and can export a single evidence ZIP.

See DEVICE_TESTING.md for a real-device test checklist.

Tests

dotnet test tests/OfflineTranscription.Tests/OfflineTranscription.Tests.csproj -c Release

Privacy

• Audio and transcription are processed locally.
• Network is used for model downloads only.
• Translate mode downloads offline translation model files on-demand.
• History and exported ZIPs may contain transcripts and audio.
• Evidence Mode can optionally include transcript text (toggle in Settings).

License

Apache License 2.0. See LICENSE and NOTICE.