وبعدها لازم أخلّص الـ presentation وأبعتها على الـ email قبل الـ deadline.
جديد
أنيسالسجلّالإعدادات
Anis for Android, app screens (redrawn from the app); anis.maslooh.com
My own example
More data, fewer mistakes
Word errors on my voice: ~29%→~13%
Same method, more data
My evaluation records: Anis versions on my own voice (99 clips), same scoring for every version; V10 = the current version
My own example
In the same league as the big models
Word errors on my voice · fewer is better
Gemini 3.1 Flash-LiteGoogle
9.2%
Anison my PC
13.2%
GPT audio miniOpenAI
18.3%
Gemini 3.5 TranscribeGoogle
18.6%
QwenCleolocal, open
19.8%
Whisper large-v3OpenAI
30.1%
VoxtralMistral
30.8%
FreeOn my deviceNo subscription
My evaluation records: same 99 clips of my voice and the same scoring for every model (16 Oct 2026; Anis = the current version)
My own example
Much faster too
Hours of audio transcribed in 1 hour · more is better
02 h4 h6 h8 h
Whisper large-v3OpenAI
~6.9 h
Gemini 3.1 Flash-LiteGoogle
~4.8 h
QwenCleolocal, open
~3.8 h
Qwen3-ASRAlibaba
~3.5 h
Anison my PC
0 h
On the phone itself: a 42-minute lecture in ~7 minutes
My evaluation records: speed measured on the same clips, Oct 2026; Anis 127× real time on my RTX 4060, Whisper large-v3 6.9×, QwenCleo 3.8×, Qwen3-ASR 3.5× (same GPU, batch 1); Gemini 3.1 Flash-Lite real-time factor 0.209 through its API
My own example
Quality beats quantity
~29%passed
Telegram lectures collected1,031 h
~29% passed the quality check
~71% did not pass
An old rule of computing
Garbage in, garbage out.
My training records: Telegram lecture packs collected and quality-gated