| --- |
| license: cc-by-4.0 |
| library_name: onnx |
| pipeline_tag: automatic-speech-recognition |
| language: |
| - ar |
| tags: |
| - quran |
| - quranic-arabic |
| - fastconformer |
| - onnx |
| - int8 |
| - streaming |
| - speech-recognition |
| base_model: mohammed/fastconformer-quran-ar |
| --- |
| |
| # FastConformer Quran — int8 ONNX (stateless export) |
|
|
| A mirror of [mohammed/fastconformer-quran-ar-onnx-int8](https://huggingface.co/mohammed/fastconformer-quran-ar-onnx-int8) |
| (**CC-BY-4.0**), packaged for the on-device recitation engine of the |
| [Voidwave Quran](https://voidwave.com/Quran) web app (runs fully in the browser via `onnxruntime-web`). |
|
|
| The base model is a Qatar Computing-style FastConformer fine-tuned on Quranic recitations |
| (NVIDIA NeMo, `tarteel-ai/everyayah`). This export is fully causal `[128, 0]` — no caches in or out — |
| so a plain stateless loop decodes it one token per encoder frame (greedy). |
|
|
| ## Files |
|
|
| | file | size | what it is | |
| |---|---|---| |
| | `encoder.int8.onnx` | ~126 MB | streaming encoder: `audio_signal [1, 80, T]` + `length [1]` → `[1, 512, T/8]` + `encoded_lengths` | |
| | `decoder.int8.onnx` | ~5 MB | **fused decoder_joint**: `encoder_outputs [1, 512, 1]`, `targets [1, 1]`, `target_length [1]`, `input_states_1/2 [1, 1, 640]` → joint logits `[1, 1, 1, 1025]` + output states | |
| | `tokens.txt` | ~13 KB | 1025 entries: 1024 BPE tokens (with diacritics) + `<blk>` at id 1024 | |
| | `meta.json` | <1 KB | upstream metadata | |
| |
| Notes: |
| |
| - The upstream `joiner.int8.onnx` is a **byte-identical duplicate** of `decoder.int8.onnx` and is omitted here. |
| - Runs on `onnxruntime-web` **≥ 1.30** (wasm kernels; earlier runtimes lack the quantized `ConvInteger` op). |
| - Input features must be NeMo-style log-mel: 80 filters, 25 ms window / 10 ms hop, `n_fft` 512, slaney mel, |
| `log(x + 2^-24)`, and **per-feature normalization is required** — despite `meta.json` claiming |
| `normalize_type: ""` (upstream metadata is wrong on this point; without normalization the decoder emits |
| an empty result). |
|
|
| ## Usage (browser, onnxruntime-web) |
|
|
| ```js |
| const ort = await import('https://cdn.jsdelivr.net/npm/onnxruntime-web@1.30.0/dist/ort.min.js'); |
| const encoder = await ort.InferenceSession.create('encoder.int8.onnx', { executionProviders: ['wasm'] }); |
| const decoder = await ort.InferenceSession.create('decoder.int8.onnx', { executionProviders: ['wasm'] }); |
| // feature extraction: NeMo-compatible mel + per-feature normalization (see notes above) |
| // decoding: encoder → per-frame greedy join with a single re-primed predictor state |
| ``` |
|
|
| ## Attribution |
|
|
| - Base model: [mohammed/fastconformer-quran-ar](https://huggingface.co/mohammed/fastconformer-quran-ar) |
| (NVIDIA NeMo FastConformer, fine-tuned on `tarteel-ai/everyayah`) |
| - Original ONNX export: [mohammed/fastconformer-quran-ar-onnx-int8](https://huggingface.co/mohammed/fastconformer-quran-ar-onnx-int8) |
| - License: **CC-BY-4.0** (see upstream) — please keep this attribution when redistributing. |
|
|
