KasuleTrevor/cdli-slam-asr-luganda-atypical-zero-shot-asr-adapted

CDLI SLAM-ASR Luganda Atypical Speech Zero-Shot ASR-Adapted Baseline

Zero-shot transfer baseline on the CDLI Luganda atypical speech dataset using the ASR-adapted SLAM-ASR checkpoint before any atypical-speech adaptation. This repository is provided to support controlled comparison against projector-only and encoder-LoRA adaptation.

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: baseline_checkpoint
• Architecture: Whisper encoder (Sunbird/asr-whisper-large-v3-salt) + linear projector + Sunflower-14B decoder; no atypical-speech adaptation; no PEFT adapters active at decode time.
• 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): 72.34%
• Normalized CER (JiWER scorer): 42.98%
• Atypical overall normalized WER: 75.52%
• Atypical overall normalized CER: 43.31%
• Atypical averaged utterance WER: 68.64%
• Atypical averaged utterance CER: 34.64%

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=false.

Additional results notes

Test subgroup breakdown: Mild 63.20% WER, Moderate 66.56%, Severe 77.25%. By disorder: Dysarthria 61.20%, Articulation Disorders 68.74%, Stuttering 70.23%, Voice disorder 81.35%. Average hyp/ref ratio was 87.85%, indicating frequent under-generation on more difficult speakers.

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.