	English	فارسی
🦀 Runtime	Pure Rust with tract	اجرای خالص Rust با tract
🌊 Mode	True cache-aware streaming	اجرای جریانی واقعی با cache
🪶 Smallest graph	138 MB INT4	کوچک‌ترین گراف INT4 با حجم ۱۳۸MB
🎧 Input	121-frame log-mel chunks	chunkهای ۱۲۱ فریمی log-mel
🔌 Dependencies	No ONNX Runtime or C++ engine	بدون ONNX Runtime یا موتور C++

Canonical repository: Reza2kn/Shenava-Koochik-v1.0-tract-streaming
PersianML mirror: PersianML/Shenava-Koochik-v1.0-tract-streaming
Parent: Reza2kn/Shenava-Koochik-v1.0

📦 Pick a graph

File	Precision	Size	Intended use
`model.int4.onnx`	Q4_0 weight-only	138 MB	Recommended edge/streaming graph.
`model.int8.onnx`	dynamic INT8 matmul	174 MB	Useful where INT8 kernels win; can be slower for batch-1 streaming.
`model.onnx`	FP32	459 MB	Reference graph.

The release’s parity checks reported byte-identical greedy transcripts for INT4 and INT8 versus FP32. The published x86 measurement for FP32 was about 83.9 ms per 1.12-second chunk (RTF about 0.075); treat this as release-specific, not a promise for other hardware.

🌊 Streaming contract

audio_signal: FP32 [1, 80, 121], unnormalized NeMo log-mel.
length: INT64 [1], the true valid frame count.
cache_last_channel: [1, 17, 70, 512].
cache_last_time: [1, 17, 512, 8].
cache_last_channel_len: INT64 [1].
Output: CTC log-probabilities plus the next cache tensors.

Feed 121-frame chunks with a shift of 112 frames and 9 frames of pre-encode overlap. Pad the first and final chunks, but pass their true lengths. Initialize cache tensors to zero, thread the returned caches into the next step, and preserve the previous CTC token across chunk boundaries when collapsing repeats. Blank id is 1024; map tokens with tokens.txt and convert SentencePiece ▁ to spaces.

🧩 tract compatibility

NeMo FastConformer cache-aware graph support was merged into tract main in sonos/tract#2441. Use a tract revision containing that change. shenava_tract_streaming.patch remains in this repository for older tract revisions and reproducibility.

persian_itn.py optionally converts spoken Persian numbers to Persian digits.

🇮🇷 خلاصهٔ فارسی

این مخزن مدل جریانی cache-aware برای موتور خالص Rust یعنی tract است. برای دستگاه‌های edge فایل INT4 پیشنهاد می‌شود. cacheها را از هر مرحله به مرحلهٔ بعد منتقل کنید و هنگام CTC collapse توکن قبلی را بین chunkها حفظ کنید.

🌌 Explore Shenava-1

🧠 Koochik source · 🦀 tract streaming · 📼 tract offline · 🌐 Browser ONNX · 🍎 Core ML

Apache-2.0.

PersianML/Shenava-Koochik-v1.0-tract-streaming

🦀🌊🎙️ Shenava Koochik v1.0 — tract streaming

✨ At a glance | معرفی سریع

📦 Pick a graph

🌊 Streaming contract

🧩 tract compatibility

🇮🇷 خلاصهٔ فارسی

🌌 Explore Shenava-1

PersianML/
Shenava-Koochik-v1.0-tract-streaming