Saboor Sarem

Software Engineer

I design and ship software end to end: native Windows applications in C# / .NET, Python tooling for local LLMs, and AVA — an offline multilingual speech-to-text app that runs entirely on the user's machine and downloads only the languages they need.

AVA

Multilingual speech-to-text for Windows: 19 offline models covering Persian, English, Russian, Chinese, Japanese, Korean, Vietnamese and dozens more. One floating orb, nothing else: tap it, speak, tap again — the text is already typed into whatever application you were writing in. The installer is only the core app; you download just the languages you actually speak.

  • 19 models
  • 100+ languages
  • Auto language
  • Pick & download
  • Offline
Download installer · 109 MB
  • Windows Download · 109 MB
  • macOS Source build
  • Linux Source build
  • Android Source build
  • iOS Source build
  • iPadOS Source build
  • HarmonyOS Scaffold

What it is

AVA is a desktop dictation app for Windows in 100+ languages, including Persian (Dari), English, Russian, Chinese, Japanese, Korean and Vietnamese. It sits on the desktop as a single floating orb, records locally, recognises speech on your own computer and types the result straight into the focused window. No audio leaves the machine.

Small core, models on demand

The 109 MB installer contains only the app and its speech runtime — no model weights and no PyTorch. After install, the Models tab lists all 19 models with their languages, size and licence; you download only the ones you speak (73 MB – 1 GB each). Removing a model frees exactly its own folder. Upgrading keeps your models and history, and offers to clean up files left over from older AVA versions.

Automatic language detection, 19 models

Auto is the default card: it identifies the language of every phrase on this PC and sends the audio to the matching downloaded model. With one model downloaded it simply uses that model; with several, a small Whisper-base detector (161 MB, fetched once) chooses between them and never transcribes.

ModelLanguagesSize
Koochik (fast)Persian459 MB
Parakeet 110M (fast)English458 MB
Parakeet 0.6B v3English + 24 European671 MB
Moonshine tinyEnglish124 MB
GigaAM v3Russian225 MB
Zipformer (small)Russian73 MB
Whisper tiny / base / small / turbo~99 languages104 / 161 / 375 / 1037 MB
SenseVoicezh, en, ja, ko, Cantonese239 MB
Dolphin40 Asian & Middle-Eastern (fa, ar, hi, ur, th, id, vi, tr …)104 MB
Fast Conformer10 European (de, en, es, fr, it, pl, ru, uk …)461 MB
Paraformer smallChinese (Mandarin)82 MB
TeleSpeechChinese dialects341 MB
FireRedASR2Chinese / English776 MB
ZipformerKorean76 MB
Zipformer (ReazonSpeech)Japanese178 MB
ZipformerVietnamese77 MB

Auto starts from Koochik (Persian), Parakeet 110M (English) and GigaAM (Russian) as soon as you download them. Every model is fetched once from Hugging Face, cached locally, and runs offline on the CPU through sherpa-onnx; after that AVA works fully offline.

Privacy & local processing

  • 100% Offline. All audio processing and transcription are executed entirely on your local machine.
  • Total Privacy. Your voice data and text are never transmitted to any cloud servers or third-party APIs.
  • Volatile Memory. Audio chunks are processed in RAM and discarded immediately after typing.
  • Downloads only. The only outbound requests are the one-time model downloads from Hugging Face and the release check against this website.

Third-party licenses

AVA runs fully offline and incorporates the following open-weights models and libraries. Model weights are downloaded on first use, not bundled, and keep their own licences. Application code is MIT.

  • Shenava-Koochik v1.0 (Persian, fast CTC) — Apache License 2.0 (apache.org)
  • NVIDIA Parakeet (English, 110M and 0.6B v3) — parakeet-tdt-0.6b-v3 · CC-BY-4.0
  • Moonshine (English) — usefulsensors/moonshine · MIT
  • GigaAM v3 (Russian) — ai-sage/GigaAM-v3 · MIT
  • Zipformer (Russian, Korean, Japanese, Vietnamese) — Apache License 2.0; Japanese weights from ReazonSpeech
  • OpenAI Whisper (tiny / base / small / turbo; base also serves as Auto's language detector) — openai/whisper-base · MIT
  • SenseVoice (zh / en / ja / ko / yue) — FunAudioLLM/SenseVoice · MIT
  • Dolphin (40 Asian & Middle-Eastern languages) — DataoceanAI/Dolphin · Apache License 2.0
  • Fast Conformer (10 European languages) — NVIDIA NeMo · CC-BY-4.0
  • Paraformer (Mandarin) — FunASR · MIT
  • TeleSpeech (Chinese dialects) — Apache License 2.0
  • FireRedASR2 (Chinese / English) — FireRedTeam/FireRedASR · Apache License 2.0
  • sherpa-onnx — k2-fsa/sherpa-onnx offline speech engine · Apache License 2.0

AVA — Your Voice. Your Words.
Built & Engineered by DANISHYARSI

01Projects

02Contact