KasuleTrevor/cdli-slam-asr-luganda-atypical-encoder-lora-alllinear-step5000

CDLI SLAM-ASR Luganda Atypical Speech All-Linear Encoder-LoRA Checkpoint (Step 5000)

All-linear encoder-LoRA atypical-speech adaptation checkpoint for SLAM-ASR on the CDLI Luganda atypical speech dataset. Low-rank adapters are applied to encoder q_proj, v_proj, k_proj, out_proj, fc1, and fc2 while the Sunflower-14B decoder remains frozen.

What this repository contains

This Hub repository stores a partial SLAM-ASR checkpoint for use with the SLAM-LLM codebase. It is not a standalone transformers checkpoint.

• Checkpoint type: encoder_lora_projector
• Architecture: Whisper encoder (Sunbird/asr-whisper-large-v3-salt) + linear projector + Sunflower-14B decoder; encoder LoRA on q_proj/v_proj/k_proj/out_proj/fc1/fc2; LLM frozen.
• Base encoder: Sunbird/asr-whisper-large-v3-salt
• Base LLM: Sunbird/Sunflower-14B
• Exported files: model.pt

Training / evaluation context

• Dataset: cdli/ugandan_luganda_nonstandard_speech_v1.0
• Evaluation split: test
• Training speakers: 36
• Validation speakers: 5
• Speaker overlap: No speaker overlap between train and validation/test

Reported metrics

• Normalized WER (JiWER scorer): 61.40%
• Normalized CER (JiWER scorer): 22.74%
• Atypical overall normalized WER: 61.57%
• Atypical overall normalized CER: 22.75%
• Atypical averaged utterance WER: 56.39%
• Atypical averaged utterance CER: 19.41%

Decode settings used for the reported metrics

Test decode used MAX_NEW_TOKENS=200, NUM_BEAMS=4, REPETITION_PENALTY=2.0, NO_REPEAT_NGRAM_SIZE=2, USE_ENCODER_PEFT=true, ENCODER_TARGET_MODULES=[q_proj,v_proj,k_proj,out_proj,fc1,fc2].

Additional results notes

Test subgroup breakdown: Mild 51.04% WER, Moderate 54.09%, Severe 65.15%. By disorder: Dysarthria 50.60%, Articulation Disorders 56.73%, Stuttering 56.88%, Voice disorder 68.91%. Average hyp/ref ratio was 98.97%, so decoding remained stable but did not outperform the narrower encoder-LoRA baseline.

Loading notes

Load through SLAM-LLM; this repository stores a partial SLAM-ASR checkpoint, not a standalone Transformers model.

Typical decode flow in this project uses:

• examples/asr_luganda/scripts/decode_luganda_sunflower.sh
• USE_ENCODER_PEFT=true for encoder-LoRA checkpoints
• matching LoRA target modules at decode time

Caveats

• This repository stores SLAM-ASR training artifacts intended for research use.
• The checkpoint must be used with the matching SLAM-LLM model code and base components.
• Results can be sensitive to decode settings and evaluation protocol.