KasuleTrevor/cdli-qwen3-asr-en-finetune-1

CDLI Qwen3-ASR 1.7B English Fine-Tune

This repository contains a Qwen3-ASR model fine-tuned from Qwen/Qwen3-ASR-1.7B 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.

Model Details

• Base model: Qwen/Qwen3-ASR-1.7B
• Fine-tuning framework: Qwen3-ASR SFT with PEFT/LoRA
• Language: English
• Checkpoint reported: checkpoint-3882

Dataset

• Dataset: cdli/ugandan_english_nonstandard_speech_v1.0
• License: cc-by-sa-4.0 (dataset card)
• Split sizes used by the source dataset card:
  • train: 5176
  • validation: 638
  • test: 1017

Evaluation artifacts in this run contain 1013 scored rows and 1013 valid normalized references.

Training Configuration

• Work root: /jupyter_kernel/qwen3_asr_cdli_en
• Base checkpoint: Qwen/Qwen3-ASR-1.7B
• Max manifest audio length: 30.0 s
• Max training audio length: 30.0 s
• Min audio length: 0.2 s
• Train batch size: 4
• Gradient accumulation steps: 2
• Effective train batch size: 8
• Learning rate: 2e-5
• Scheduler: cosine

Evaluation

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

Corpus Metrics

• Raw WER: 29.55%
• Raw CER: 16.15%
• Normalized WER: 24.06%
• Normalized CER: 14.93%

Average Utterance Metrics

• Average normalized utterance WER (capped at 1.0): 22.72%
• Average normalized utterance CER (capped at 1.0): 14.35%

Usage

from qwen_asr import Qwen3ASRModel

model = Qwen3ASRModel.from_pretrained('KasuleTrevor/cdli-qwen3-asr-en-finetune-1')
predictions = model.transcribe(['path/to/audio.wav'])
print(predictions[0].text if hasattr(predictions[0], 'text') else predictions[0])

Files

• model.safetensors: fine-tuned model weights
• tokenizer.json, tokenizer_config.json, preprocessor_config.json, chat_template.json
• results/checkpoint-3882/test_predictions.csv
• results/checkpoint-3882/test_predictions.jsonl
• results/checkpoint-3882/test_predictions_scored.csv
• results/checkpoint-3882/test_predictions_scored.jsonl
• results/checkpoint-3882/test_predictions_grouped_analysis.csv

Notes

• Results are stored in a checkpoint-scoped folder under results/.
• Normalized metrics use transcript normalization to reduce punctuation, casing, and formatting noise during evaluation.
• Access to the source dataset is gated. Review dataset terms before requesting access.