Nemo Arabic STT Diacritized

--- license: cc-by-4.0 language: - ar tags: - automatic-speech-recognition - arabic - diacritization - tashkeel - speech pipeline_tag: automatic-speech-recognition --- # Nemo-Arabic-STT-Diacritized An Arabic speech-to-text pipeline that produces fully diacritized (tashkeel) transcripts. It combines two independently developed models in sequence; it is a packaged inference pipeline, not a single jointly trained architecture. ## Pipeline 1. **Speech recognition**: [NVIDIA NeMo FastConformer Hybrid](https://huggingface.co/nvidia/stt_ar_fastconformer_hybrid_large_pcd_v1.0) (`stt_ar_fastconformer_hybrid_large_pcd_v1.0`) transcribes Arabic speech to plain, undiacritized text. 2. **Diacritization**: [CATT](https://github.com/abjadai/catt) (Character-based Arabic Tashkeel Transformer) adds tashkeel to the transcript. Audio in, diacritized Arabic text out. Neither checkpoint was retrained or fine-tuned for this repository. ## Files | File | Description | Source | |---|---|---| | `stt_ar_fastconformer_hybrid_large_pcd_v1.0.nemo` | ASR checkpoint, unmodified | NVIDIA, CC-BY-4.0 | | `best_ed_mlm_ns_epoch_178.pt` | Diacritizer checkpoint, unmodified | abjadai/CATT, Apache-2.0 | | `diacritize.py`, `catt/` | Diacritizer inference code (vendored from CATT) | abjadai/CATT, Apache-2.0 | | `pipeline.py` | `DiacritizedASR` — loads both models once, audio in / diacritized text out in a single call | This repository | | `server.py` | Reference FastAPI server implementing the full pipeline | This repository | ## Usage ```python from pipeline import DiacritizedASR model = DiacritizedASR( nemo_path="stt_ar_fastconformer_hybrid_large_pcd_v1.0.nemo", catt_ckpt="best_ed_mlm_ns_epoch_178.pt", ) diacritized = model.transcribe("audio.wav") ``` Both models load once at construction; each `.transcribe()` call runs ASR followed immediately by diacritization in the same process. For the two steps individually: ```python import nemo.collections.asr as nemo_asr from diacritize import Diacritizer asr_model = nemo_asr.models.EncDecHybridRNNTCTCBPEModel.restore_from( "stt_ar_fastconformer_hybrid_large_pcd_v1.0.nemo" ) diacritizer = Diacritizer(ckpt="best_ed_mlm_ns_epoch_178.pt") text = asr_model.transcribe(["audio.wav"])[0].text diacritized = diacritizer.diacritize_texts([text])[0] ``` Or run `server.py` directly for an HTTP API (`POST /transcribe`, `GET /health`). ## Exam...