AvaSanj-v1

Persian repr1 automatic speech recognizer (AvaSanj v1) fine-tuned from facebook/wav2vec2-xls-r-300m on 8,437 approved Persian audio clips.

Outputs phonemes in the shared repr1 alphabet (A C Q S Z … + lowercase IPA-style phonemes) — the same canonical alphabet used by the Negara G2P models, making this model a drop-in listener for grapheme-to-phoneme adjudication.

Provenance

• Original model: facebook/wav2vec2-xls-r-300m (Apache-2.0), fine-tuned for CTC.
• Data: 8,437 approved (lattice-supported) Persian clips; target labels are the audio_target_phonemes_vnext repr1 label field.
• Role: used as the OOF (out-of-fold) listener in the AvaSanj data pipeline — each fold scores only voices it never trained on — to arbitrate G2P disagreements at margin ≥ 0.1 (~96.3% precision / 95.9% recall on the corruption gate).
• Selection: best step 800, best validation CER 0.0479.

Model

• Architecture: Wav2Vec2ForCTC, base wav2vec2-xls-r-300m (~300 M params).
• Vocabulary: repr1 (see vocabulary.json) — 30-token inventory incl. <blank> + " ".
• Full eval: report.json.

Usage

from transformers import Wav2Vec2ForCTC
model = Wav2Vec2ForCTC.from_pretrained("Reza2kn/AvaSanj-v1-audio8437")

Note: load with the local vocabulary.json for the repr1 tokenizer mapping.

Files

• config.json, model.safetensors — HF wav2vec2 checkpoint (best-model/)
• vocabulary.json — repr1 vocab
• report.json — full eval report (baseline, corruption gate, OOF fold stats)