cdli-qwen3-asr-lg-atypical-stage3-0p6b – AI Model by KasuleTrevor | AlphaNeural AI

You can deploy this model and start earning money today!

KasuleTrevor/
cdli-qwen3-asr-lg-atypical-stage3-0p6b

like

0

automatic-speech-recognition

atypical-speech

cdli/ugandan_luganda_nonstandard_speech_v1.0

KasuleTrevor/cdli-qwen3-asr-lg-textpre-0p6b-v2

endpoints_compatible

Views

No views yet

CDLI Qwen3-ASR Luganda Atypical Speech Fine-tune

This repo contains the selected checkpoint checkpoint-2000 from the LG-QWEN3-ASR-ATYPICAL-STAGE3-0P6B run.

Training Setup

Base model: KasuleTrevor/cdli-qwen3-asr-lg-textpre-0p6b-v2
Dataset: cdli/ugandan_luganda_nonstandard_speech_v1.0
Split files: train_cleaned.tsv, validation_cleaned.tsv, test_cleaned.tsv
Training language tag: Luganda
Forced inference language: disabled
Epochs: 3
Batch size: 4
Gradient accumulation: 2
Learning rate: 0.0001
Scheduler: cosine
Warmup ratio: 0.03
Save steps: 500
Selected checkpoint: checkpoint-2000
Selection reason: best validation normalized WER (0.656588, CER 0.336131)

Final Test Metrics

Corpus WER (normalized): 0.734365
Corpus CER (normalized): 0.417267
Average utterance WER (normalized): 0.514141
Average utterance CER (normalized): 0.249370

Checkpoint Selection Evidence

checkpoint	wer_normalized	cer_normalized	avg_wer_normalized	avg_cer_normalized	eval_loss
checkpoint-2000	0.6565882352941177	0.3361308034473172	0.590703736636411	0.310064505120198	1.0837534666061401
checkpoint-2211	0.6711764705882353	0.3523943564081179	0.5916740338341632	0.30936556868573584	1.0837534666061401
checkpoint-1500	0.8458823529411764	0.512892688351404	0.5885575362296198	0.3056259208190241	1.0516055822372437
checkpoint-1000	0.9490588235294117	0.5825166805671392	0.6633167229642679	0.3610897035439228	1.1547762155532837
checkpoint-500	1.016470588235294	0.7321726438698916	0.738168804683814	0.4641290685476656	1.3621718883514404

Severity Breakdown

severity_speech_impairment	n_samples	n_speakers	mean_wer	mean_cer	median_wer	median_cer
Severe (frequent breakdowns)	315	3	0.6388	0.3523	0.6667	0.3025
Moderate (requires effort to understand)	347	3	0.5078	0.2348	0.5	0.1647
Mild (easily understood with minimal effort)	365	3	0.4126	0.1744	0.3333	0.1061

Artifacts

Result folder: results/checkpoint-2000/
Includes checkpoint validation summaries, final test predictions, scored outputs, and grouped metadata analyses.