Talqin streaming model v1.0 (Warsh)
ุชูููู: ูู ูุฐุฌ ุตุบูุฑ ูุณู ุน ุชูุงูุฉ ุงููุฑุขู ุงููุฑูู ุจุฑูุงูุฉ ูุฑุด ุนู ูุงูุน ู ุจุงุดุฑุฉู ุนูู ุงููุงุชูุ ููุชุงุจุน ุงููุงุฑุฆ ููู ุฉู ููู ุฉ ููุนุฑู ู ูุถุนู ู ู ุงูู ุตุญู.
A small model that hears a Quran recitation (riwayah of Warsh from Nafi') as it comes, on the reciter's own phone. The Talqin app uses it to follow the reciter word by word and find the ayah, with no server.
- What it writes: the recitation's sounds (60 symbols,
tokens.txt), one every 40 ms, about 0.4 s after they are said. Not words: the Talqin app places these sounds in the Quran. - Model: a causal (streaming) Zipformer with CTC, 22.3M parameters, trained from scratch on 1,511 hours of 59
reciters, with noise, phone-line and speed effects.
model.int8.onnx(25 MB) hears 0.32 s slices, 8-bit. - Runs with: sherpa-onnx (
OnlineRecognizer,zipformer2Ctc), on Android, iPhone, computers and in the browser (WebAssembly). It runs 2โ15 times faster than real time on phones. - Accuracy: 1.8% sound error on 7 reciters it never heard (8.45 h); in a simulated home room (echo, noise, phone microphone) 9.1%. The app's search places 98.3% of phrases right with it.
Files
| File | What it is |
|---|---|
model.int8.onnx |
The model (0.32 s slices, 8-bit) |
tokens.txt |
Its 60 sounds |
manifest.json |
Version, sizes and SHA-256 of the files (the app checks them) |
Use
import sherpa_onnx
rec = sherpa_onnx.OnlineRecognizer.from_zipformer2_ctc(tokens="tokens.txt", model="model.int8.onnx", num_threads=1)
Limits
It follows the recitation; it is not a judge of mistakes (it heard the same Quran text many times in training, so it partly knows it by heart). Talqin checks mistakes with a larger model on its server.
ูุฐุง ุงููู ูุฐุฌ ูุนู ู ุจุงูุฐูุงุก ุงูุงุตุทูุงุนู ููุฏ ูุฎุทุฆุ ููู ููุฅุนุงูุฉ ููุท ููุง ูุฌุจ ุงูุงุนุชู ุงุฏ ุนููู 100% ุฃู ุชูุฏูู ู ุจุฏููุงู ุนู ู ุนููู ู ุชูู ูู ุฌุงุฒ.
License
Apache License 2.0. Made by Talqin (https://github.com/TalqinCore).