Shenava 1.0

Central release hub for Shenava 1.0, an open streaming Persian ASR and captioning stack spanning trained models, deployable runtimes, datasets, semantic evaluation, leaderboard tooling, and public applications.

Start Here

Area	Link	What it is for
Release collection	🤗 Shenava 1.0 Collection	Models, exports, datasets, benchmark artifacts, and demos cited by the paper.
SLT paper	OpenReview submission	Shenava / VisualEars paper submitted to SLT.
Persian app	🔗 Shenava	Persian real-time captioning app.
English app	🔗 VisualEars	English real-time captioning app.
Leaderboard	Persian ASR Triple Threat	Public Persian ASR benchmark and ranking table.

Implementation Repositories

Repository	Component	Notes
shenava-ctc-beam	Rust CTC beam decoder	Static-3669 decoding path and Persian word-guide integration.
shenava-esp32s3	ESP32-S3 runtime	Bare-device streaming ASR demo path for the 6.9M Rizeh-Pizeh model.
Vosk-Rust	Rust Vosk/Kaldi-style decoder	Rust port released for users who need the Vosk-style decoding stack.

Hugging Face Resources

Model cards

Model	Scale	Deployment cards
Shenava Koochik v1.0	114M	ONNX FP16 · CoreML FP16 · iOS 15 CoreML · sherpa-onnx · tract streaming
Shenava Rizeh v1.0	32M	ONNX FP16 · CoreML FP16 · iOS 15 CoreML · sherpa-onnx · tract streaming
Shenava Rizeh-Pizeh v1.0	6.9M	ONNX FP16 · CoreML FP16 · iOS 15 CoreML · sherpa-onnx · tract streaming
ShenavaSanj v1.0	0.2B	Persian word-importance model used by S3 / Semantic WER.

Dataset cards

Dataset	Purpose
visualears-persian-asr-16k	Main public 16 kHz Persian ASR audio/text release.
visualears-full-pipeline-state	Phase A/B pipeline state, audit manifests, curriculum accounting, and robustness assets.
visualears-pseudo-clean-train	Pseudo-clean training view used by the de-poisoning pipeline.
persian-asr-audio-text-2.69M-chizzled	Phase A-scale filtered audio/text corpus.
persian-asr-relabeled-gemini	Gemini-relabeled Persian ASR subset.
visualears-115m-pref350-curriculum-relabels	Koochik curriculum and relabel artifacts.
visualears-golden-6669	Golden6669 held-out benchmark.
fleurs-fa-benchmark	FLEURS-fa benchmark packaging used by the leaderboard.
visualears-benchmark-269-gold	269-record noisy gold benchmark.
visualears-active-learning-669-audio	Active-learning correction audio.
visualears-al-phaseB2-audio	Phase B2 active-learning audio.
visualears-al-114m-audio	Koochik 114M active-learning set.
visualears-al-32m-audio · visualears-al-32mv2-audio	Rizeh 32M active-learning sets.
visualears-hardword-sentences	Semantic and keyword stress sentences, including audio configurations.
telephooney-trainability-splits · telephooney-raw-audio	Telephooney training splits and raw source audio.
golha-persian-spoken · golha-asr-gold-69	Golha spoken-word corpus and its gold evaluation set.
ganjoor-recitations-chunked	Training-ready Persian poetry recitation chunks.

Demos and evaluation

Resource	Type	Notes
Shenava 1.0 Collection	Collection	One place for the full Shenava 1.0 release family.
Persian ASR Triple Threat	Space	Public leaderboard with WER, CER, S3, ESS, and aggregate ranking.
WebGPU Captioning Demo	Space	Browser captioning demo path.

Dataset Scale

Dataset or pool	Rows / items	Hours	Notes
visualears-persian-asr-16k	3,927,733 rows	7,771.12 h	Public HF release; 509.0 GB parquet/audio data.
Audited Phase A/B curriculum pool	3,803,064 rows	7,685.12 h	Published pipeline state and audited training curriculum accounting.
Relabel pool	454,675 rows	690.00 h	Prior relabel pool and its published relabeled view.
Filtered relabel / curriculum artifacts	231,959 rows	412.31 h	Filtered subset used in release accounting.
Targeted noisy-speech batch	-	269.00 h	Robustness batch documented in the published pipeline state.
Environment-event and hard-word assets	6,969 assets	-	Non-speech/environment and semantic stress assets.
Telephooney v0 candidate video pool	3,114 videos	969.02 h	Candidate video pool used for source discovery and filtering.

Release Contents

Area	Contents
Models	Shenava Koochik 114M, Rizeh 32M, and Rizeh-Pizeh 6.9M, plus their platform exports linked above.
Datasets	Main Persian ASR release, Phase A/B pipeline state, active-learning sets, benchmark datasets, and release audit artifacts linked above.
Evaluation	ShenavaSanj, Semantic WER (S3), essential-word error, and the Triple Threat leaderboard.
Runtimes	Rust-native CTC beam decoding, tract streaming exports, CoreML, ONNX, LiteRT, WebGPU/WASM, and embedded ESP32-S3 artifacts.
Applications	Shenava for Persian captioning and VisualEars for English captioning across desktop, mobile, browser, and overlay use cases.

Licensing

Scope	License policy
Project-owned code, cards, benchmark tooling, and derived artifacts	Apache-2.0 unless a repository card states otherwise.
Third-party-derived audio sources	Original source licenses apply; dataset cards are the source of truth for per-source rights.
Non-permissive or incompatible sources	Isolated or excluded from Apache-2.0 claims.

License

Licensed under the Apache License 2.0.