CoolFace
Modelpublic

KasuleTrevor/cdli-slam-asr-english-atypical-meusli-v1-projector-only-e3s208

sourceHugging Faceotherupdated 6mo agoView on Hugging Face
0likes
Model Card

CDLI SLAM-ASR English Atypical Speech MEUSLI v1 Projector-Only Checkpoint (Epoch 3 Step 208)

Projector-only atypical-speech adaptation checkpoint for SLAM-ASR on the CDLI Ugandan English atypical speech dataset. This run starts from the SpeechTek multilingual speech LLM linear projector v1 stack and fine-tunes only the linear projector while keeping the Whisper encoder and EuroLLM decoder 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: projector_only
  • —Architecture: Whisper-large-v3-turbo encoder + linear projector (mEUltilingualspeechllmlinearprojectorv1 initialization) + EuroLLM-1.7B decoder; encoder frozen; LLM frozen; no prompt; no PEFT adapters during training or decode.
  • —Base encoder: openai/whisper-large-v3-turbo
  • —Base LLM: utter-project/EuroLLM-1.7B
  • —Exported files: model.pt

Training / evaluation context

  • —Dataset: cdli/ugandan_english_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): 29.38%
  • —Normalized CER (JiWER scorer): 19.79%
  • —Atypical overall normalized WER: 29.67%
  • —Atypical overall normalized CER: 19.81%
  • —Atypical averaged utterance WER: 28.05%
  • —Atypical averaged utterance CER: 18.80%

Decode settings used for the reported metrics

Decode used the English MEUSLI no-prompt configuration with MAXNEWTOKENS=200, NUMBEAMS=4, REPETITIONPENALTY=2.0, NOREPEATNGRAMSIZE=2, and USELLM_PEFT=false.

Additional results notes

This checkpoint improved the English zero-shot MEUSLI v1 baseline from 41.59% to 28.05% averaged utterance WER. Best speakers were UG014 (14.36%), UG021 (20.06%), and UG042 (24.16%). Hardest groups remained voice disorder (39.17%) and acquired hearing impairment (39.17%).

Loading notes

Load through SLAM-LLM with the exact English MEUSLI/OpenAI stack: openai/whisper-large-v3-turbo encoder, utter-project/EuroLLM-1.7B decoder, and no prompt template. This repository stores a partial SLAM-ASR checkpoint, not a standalone Transformers model.

Typical decode flow in this project uses project-specific wrappers such as:

  • —examples/asr_luganda/scripts/decode_luganda_sunflower.sh for the Sunflower/Luganda stack
  • —examples/asr_luganda/scripts/decode_english_meusli_openai.sh for the English MEUSLI/OpenAI stack
  • —matching PEFT settings at decode time when adapters are part of the checkpoint

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.