Kimi-Audio-7B-SAP-EP

The best model of the paper (7.77% WER on SAP). Kimi-Audio-7B-Instruct fine-tuned to first name the speaker's condition and then transcribe, in the same output. No information about the speaker is needed at inference time.

Paper: Etiology-Aware Speech Language Models for Dysarthric Speech Recognition (Interspeech 2026) Paper | Code

The four models

All four are Kimi-Audio-7B-Instruct fine-tuned on the same data. They differ only in how the speaker's condition (etiology) is used.

WER in %, lower is better. SemScore higher is better. TORGO is zero-shot (no TORGO training).

Input and output

Input: the prompt Please transcribe the spoken content into written text. and the audio.

Output:

Etiology: <ALS | Parkinson's Disease | Stroke | Down Syndrome | Cerebral Palsy>
Transcript: <transcript>

Keep the text after Transcript: as the transcription.

Usage

git clone --recursive https://github.com/MorenoLaQuatra/dysarthric-asr
cd dysarthric-asr
pip install -r requirements/kimi.txt -r kimi/third_party/Kimi-Audio/requirements.txt
bash kimi/scripts/setup.sh
source kimi/scripts/env.sh

from kimia_infer.api.kimia import KimiAudio

model = KimiAudio(model_path="morenolq/Kimi-Audio-7B-SAP-EP", load_detokenizer=False)
messages = [
    {"role": "user", "message_type": "text", "content": "Please transcribe the spoken content into written text."},
    {"role": "user", "message_type": "audio", "content": "speech.wav"},
]
_, text = model.generate(messages, output_type="text",
                         text_temperature=0.0, text_top_k=5,
                         audio_temperature=0.8, audio_top_k=10)
print(text)

To keep only the transcription: transcript = text.split("Transcript:", 1)[-1].strip()

For batch inference and the paper evaluation (SAP and TORGO), see the code repository.

Results per condition

Etiology prediction accuracy on SAP: 71.7%.

Training

• Data: Speech Accessibility Project (SAP), release v31-05-2025. 240,047 utterances (547 h) from speakers with ALS, Parkinson's disease, stroke, Down syndrome and cerebral palsy.
• Method: LoRA (rank 16, alpha 32) on the language model and on the Whisper encoder attention. 3 epochs, learning rate 5e-5, global batch 64. Checkpoint chosen by validation loss.
• This repository holds the merged weights (fp32) exactly as evaluated in the paper.

Limitations

• For research on English dysarthric speech recognition.
• The predicted etiology is a helper for transcription, not a diagnosis. Do not use it for clinical decisions.
• Trained on five conditions. Other conditions, languages or recording setups are not tested beyond TORGO.
• No audio or transcripts from SAP are included in this repository.

Citation

If you use this model, please cite the paper:

@inproceedings{laquatra26_interspeech,
  title     = {{Etiology-Aware Speech Language Models for Dysarthric Speech Recognition}},
  author    = {Moreno {La Quatra} and Alkis Koudounas and Valerio Mario Salerno and Sabato Marco Siniscalchi},
  year      = {2026},
  booktitle = {{Interspeech 2026}},
  pages     = {5033--5037},
  doi       = {10.21437/Interspeech.2026-1773},
  issn      = {2958-1796},
}

Acknowledgements

Trained on data from the Speech Accessibility Project (Hasegawa-Johnson et al., 2024). Base model: Kimi-Audio (Ding et al., 2025). License: MIT, as the base model.