Modelpublic

Peacockery/parakeet-ctc-109m-farsi

sourceHugging Facecc-by-4.0updated 4mo agoView on Hugging Face

0likes158downloads

Model Card

parakeet-ctc-109m-farsi

A Farsi fine-tune of NVIDIA's Parakeet CTC 109M (the CTC branch of parakeet-tdt_ctc-110m), trained with a custom Farsi SentencePiece tokenizer on 844 hours of quality-classified Farsi speech. It is the best Parakeet-family model from this project; a later variant trained on a stricter 224-hour subset regressed and was retired.

Files

—model.nemo: the complete NeMo checkpoint (weights + tokenizer + config)

Training data

844 hours of Farsi speech: Common Voice, Thomcles, Mana TTS, Neyshekar, FLEURS, WorldSpeech, and Farsi YouTube, with web-scraped portions machine-labeled and filtered by a Scribe quality classifier. Clips capped at 400 characters and 60 chars/sec to drop misaligned transcripts.

Benchmarks

Test splits, corpus-level jiwer after shared normalization, re-scored 2026-05-28 (WER %):

split	WER	CER
FLEURS	14.94	4.15
Common Voice 25	23.19	6.27
Mana TTS	28.53	5.07
Neyshekar	27.87	5.48
WorldSpeech	40.12	20.51
YouTube	34.55	13.72

For a stronger Farsi model from the same project see Peacockery/omni-ctc-300m-farsi, which wins on every split. This checkpoint is the lighter and faster option (109M parameters, ~350x realtime on the benchmark hardware).

Usage

python

import nemo.collections.asr as nemo_asr
model = nemo_asr.models.EncDecCTCModelBPE.restore_from("model.nemo")
texts = model.transcribe(["audio.wav"])