KasuleTrevor/cdli-parakeet-en-ug-ke-1.1b

CDLI Parakeet TDT 1.1B English Fine-Tune (Uganda + Kenya)

This repository contains a NeMo ASR model fine-tuned from nvidia/parakeet-tdt-1.1b on a merged atypical-English training mix from Uganda and Kenya.

The setup mirrors the mixed-country 0.6B notebook, but swaps in the 1.1B checkpoint and keeps the same evaluation pattern: Uganda-only validation for selection, then separate held-out reporting for Uganda and Kenya.

Model Details

• Base model: nvidia/parakeet-tdt-1.1b
• Fine-tuning framework: NVIDIA NeMo
• Language: English
• Acoustic model family: FastConformer-TDT / RNNT-BPE
• Output text: lower-case English transcription with standard ASR normalization

Datasets

• Uganda: cdli/ugandan_english_nonstandard_speech_v1.0
• Kenya: cdli/kenyan_english_nonstandard_speech_v1.0
• License: cc-by-sa-4.0
• Audio sampling rate: 16 kHz

Split sizes used in this run:

• Uganda train: 5175
• Uganda validation: 638
• Uganda test: 1016
• Kenya train: 4374
• Kenya validation: 542
• Kenya test: 928
• Merged train total: 9549 rows, 56.46 hours
• Validation selection set: Uganda only, 638 rows, 4.33 hours

Training Configuration

• Work root: /jupyter_kernel/parakeet_cdli_en_ug_ke
• Train mix: Uganda + Kenya
• Primary validation country: Uganda
• Max manifest audio length: 40.0 s
• Max training audio length: 40.0 s
• Min audio length: 0.2 s
• Train batch size: 8
• Eval batch size: 8
• Gradient accumulation steps: 8
• Effective train batch size: 64
• Learning rate: 5e-5
• Weight decay: 1e-3
• Warmup steps: 100
• Scheduler: CosineAnnealing
• Max steps: 20000
• Validation interval: 200 steps
• Early stopping patience: 12
• Precision: bf16-mixed when supported, otherwise mixed precision fallback

Evaluation

Evaluation was run separately on the held-out Uganda and Kenya test splits using both raw transcript comparison and normalized transcript comparison.

Uganda Test Set

• Raw WER: 33.14%
• Raw CER: 15.53%
• Normalized WER: 21.59%
• Normalized CER: 12.92%
• Average normalized utterance WER (capped at 1.0): 20.86%
• Average normalized utterance CER (capped at 1.0): 12.88%

Kenya Test Set

• Raw WER: 33.62%
• Raw CER: 11.92%
• Normalized WER: 12.45%
• Normalized CER: 7.59%
• Average normalized utterance WER (capped at 1.0): 12.48%
• Average normalized utterance CER (capped at 1.0): 7.63%

Uganda and Kenya Comparison

The same shared 1.1B model performs much better on the Kenyan test split than on the Ugandan test split.

• Kenya vs Uganda normalized WER: 12.45% vs 21.59%
• Kenya vs Uganda normalized CER: 7.59% vs 12.92%
• Absolute normalized WER gap: 9.14 points in favor of Kenya

Relative to the mixed-country 0.6B run (KasuleTrevor/cdli-parakeet-en-ug-ke), the 1.1B model is uneven:

• Uganda normalized WER: 21.59% vs 21.33% for 0.6B
• Kenya normalized WER: 12.45% vs 14.56% for 0.6B

So the larger model did not improve Uganda, but it improved Kenya materially by 2.11 absolute normalized WER points.

Interpretation

This run is not a clean across-the-board win over the 0.6B model. The main benefit is stronger Kenyan generalization, while Uganda remains roughly flat to slightly worse.

That matters for model selection:

• If Uganda remains the primary deployment target, the 0.6B mixed-country run is still competitive.
• If the goal is broader East African English transfer, the 1.1B run is more interesting because of the Kenya gain.

Usage

from nemo.collections.asr.models import ASRModel

model = ASRModel.from_pretrained("KasuleTrevor/cdli-parakeet-en-ug-ke-1.1b")
predictions = model.transcribe(["path/to/audio.wav"])
print(predictions[0].text if hasattr(predictions[0], "text") else predictions[0])

Files

• EN-PARAKEET-TDT-UG-KE-1.1b.nemo: exported NeMo checkpoint
• results/EN-PARAKEET-TDT-UG-KE-1.1b/: uploaded result tables and breakdown CSVs
• train_mix_summary_ug_ke.json
• country_summary_ug_ke.csv
• test_predictions_scored_uganda.csv
• test_predictions_scored_kenya.csv
• severity_breakdown_combined_ug_ke.csv
• disorder_breakdown_combined_ug_ke.csv
• etiology_breakdown_combined_ug_ke.csv

Notes

• Validation for early stopping and checkpoint selection was Uganda-only, even though training used both Uganda and Kenya.
• Source datasets are gated. Review the dataset terms before requesting access.
• Result artifacts for both countries were uploaded to the Hub under results/EN-PARAKEET-TDT-UG-KE-1.1b/.