KasuleTrevor/cdli-parakeet-en-finetune

CDLI Parakeet TDT 0.6B English Fine-Tune

This repository contains a NeMo ASR model fine-tuned from nvidia/parakeet-tdt-0.6b-v2 on the gated cdli/ugandan_english_nonstandard_speech_v1.0 dataset.

The task is English automatic speech recognition for atypical or non-standard speech from Ugandan speakers, including dysarthric speech. The dataset is part of the CDLI research collection and requires access approval.

Model Details

• Base model: nvidia/parakeet-tdt-0.6b-v2
• Fine-tuning framework: NVIDIA NeMo
• Language: English
• Acoustic model family: FastConformer-TDT / RNNT-BPE
• Output text: lower-case English transcription with standard ASR normalization

Dataset

• Dataset: cdli/ugandan_english_nonstandard_speech_v1.0
• License: cc-by-sa-4.0
• Split sizes used:
  • train: 5176
  • validation: 638
  • test: 1017
• Audio sampling rate: 16 kHz

The model was fine-tuned and evaluated on the CDLI Ugandan English non-standard speech corpus, which includes speaker metadata such as severity of speech impairment, age, gender, type of non-standard speech, and etiology.

Training Configuration

• Work root: /jupyter_kernel/parakeet_cdli_en
• Max manifest audio length: 30.0 s
• Max training audio length: 30.0 s
• Min audio length: 0.2 s
• Train batch size: 8
• Eval batch size: 8
• Gradient accumulation steps: 2
• Effective train batch size: 16
• Learning rate: 1e-4
• Weight decay: 1e-3
• Warmup steps: 100
• Scheduler: CosineAnnealing
• Training steps: 2000
• Precision: bf16-mixed when supported, otherwise mixed precision fallback

Evaluation

Evaluation was run on the held-out test split using both raw transcript comparison and normalized transcript comparison.

Corpus Metrics

• Raw WER: 27.67%
• Raw CER: 14.77%
• Normalized WER: 21.72%
• Normalized CER: 13.46%

Average Utterance Metrics

• Average normalized utterance WER (capped at 1.0): 20.98%
• Average normalized utterance CER (capped at 1.0): 13.36%

Usage

from nemo.collections.asr.models import ASRModel

model = ASRModel.from_pretrained("KasuleTrevor/cdli-parakeet-en-finetune")
predictions = model.transcribe(["path/to/audio.wav"])
print(predictions[0].text if hasattr(predictions[0], "text") else predictions[0])

Files

• EN-PARAKEET-TDT-F1.nemo: exported NeMo checkpoint
• checkpoints/: intermediate training checkpoints

Notes

• This model card reports the best-performing English Parakeet result obtained in this project between the 0.6B and 1.1B runs.
• Normalized metrics use transcript normalization to reduce punctuation, casing, and formatting noise during evaluation.
• Access to the source dataset is gated. Users should review the dataset terms before requesting access.