---
library_name: transformers
pipeline_tag: automatic-speech-recognition
language:
- en
base_model: openai/whisper-small
tags:
- automatic-speech-recognition
- whisper
- child-speech
- model-merging
- compositional-domain-adaptation
- 04_robustness_synthetic_small
- headwise-selective-attention
---

# Whisper Small HSA Merge: all TTS + Script 0-2

This repository contains a minimal Hugging Face Transformers checkpoint for
the manuscript **Compositional Domain Adaptation for Automatic Speech Recognition with Headwise Selective Attention Merging**.

## Model Details

- **Model type:** Whisper sequence-to-sequence ASR model
- **Base model:** `openai/whisper-small`
- **Release group:** Synthetic-data robustness
- **Checkpoint kind:** Headwise Selective Attention (HSA) merged checkpoint
- **Manuscript role:** Synthetic-data merge
- **Source artifact:** `04_robustness_synthetic_small/all_tts_hsa_merge`

## Method Context

This is a structured model merge for compositional domain adaptation. It composes task-specific adaptations without additional retraining by restricting parameter arithmetic to salient attention heads where the adaptations are concentrated.

Training/adaptation context: Synthetic-data composition: TTS-derived speech and scripted 0-2 adaptations.

The broader manuscript studies whether speech foundation model adaptations
for different distribution shifts, such as acoustic condition, speaking style,
speaker population, and dialect, can be recombined for low-resource and
intersectional ASR without direct joint-supervision data.

## Intended Use

Use this checkpoint to reproduce or extend the paper's ASR model-merging
experiments. It is intended for research on child ASR, compositional domain
adaptation, robustness, cross-corpus transfer, dialectal variation, and scaling
behavior across Whisper model sizes.

## How To Load

```python
from transformers import WhisperForConditionalGeneration, WhisperProcessor

model_id = "balaji1312/whisper_small_hsa_all_tts_script_0_2"
processor = WhisperProcessor.from_pretrained(model_id)
model = WhisperForConditionalGeneration.from_pretrained(model_id)
```

For local use before upload:

```python
from pathlib import Path
from transformers import WhisperForConditionalGeneration, WhisperProcessor

model_dir = Path("final_release_models") / "04_robustness_synthetic_small" / "whisper_small_hsa_all_tts_script_0_2"
processor = WhisperProcessor.from_pretrained(model_dir)
model = WhisperForConditionalGeneration.from_pretrained(model_dir)
```

## Release Files

This model card was generated for the curated release tree. The model-loading
payload consists of:

`config.json, generation_config.json, preprocessor_config.json, tokenizer_config.json, vocab.json, merges.txt, normalizer.json, special_tokens_map.json, added_tokens.json, model.safetensors`

Training state, optimizer state, decode logs, hypotheses, references, and
intermediate experiment outputs were intentionally omitted.

## Limitations

The checkpoint is released for research reproducibility. Results outside the
paper's child ASR, robustness, cross-corpus, dialectal, and scaling-law
settings are not characterized here. Reproducing WER numbers requires the
manuscript evaluation pipeline and authorized access to the relevant speech
corpora; no evaluation audio or transcripts are redistributed in this model
folder.

## Citation

If you use this checkpoint, please cite the manuscript:

```bibtex
@article{shankara2026compositional,
  title = {Compositional Domain Adaptation for Automatic Speech Recognition with Headwise Selective Attention Merging},
  author = {Shankara, Natarajan Balaji and Wang, Zilai and Eren, Eray and Alwan, Abeer},
  year = {2026},
  note = {Manuscript submitted to Computer Speech & Language}
}
```