Skip to content
Audio

Whisper-large-v3-turbo Korean Dysarthria LoRA

Adapter for Impaired Korean Speech Recognition

2 min readupdated

This lightweight PEFT LoRA module (a small add-on on top of the base model) was trained on a Korean corpus combining audio from speakers with dysarthria (articulation disorders) and typical speech. The base openai/whisper-large-v3-turbo model often struggles with non-standard pronunciation or "forgets" standard speech patterns when fine-tuned narrowly. Joint training solves this problem, improving recognition accuracy for both speakers with speech impairments and everyday users. It is a solid fit for developers building voice assistants, inclusive communication tools, and transcription services.

Key features

  • Optimized for Korean speech affected by motor and neurological articulation disorders.
  • Cuts Character Error Rate (CER) by 86.7% on dysarthria benchmarks.
  • Boosts accuracy on typical speech as well, without degrading baseline performance.
  • Trains just 6.44% of the base model's parameters (around 55.7M trainable weights).

What you can do

  • Plug the adapter into your audio pipeline to transcribe challenging or slurred speech accurately.
  • Build accessible tools and captioning systems for people with speech impairments.
  • Merge the weights (merge_and_unload) into the base Whisper model for faster production inference.
  • Test transcription on Korean lectures and audio recordings using the transformers Python library.

Example

import torch
import torchaudio
from transformers import WhisperForConditionalGeneration, WhisperProcessor
from peft import PeftModel

base_model_id = "openai/whisper-large-v3-turbo"
adapter_id = "PL-HFT/whisper-large-v3-turbo-ko-dysarthria-lora"

device = "cuda" if torch.cuda.is_available() else "cpu"
dtype = torch.float16 if device == "cuda" else torch.float32

processor = WhisperProcessor.from_pretrained(base_model_id, language="Korean", task="transcribe")
base_model = WhisperForConditionalGeneration.from_pretrained(base_model_id, torch_dtype=dtype)
model = PeftModel.from_pretrained(base_model, adapter_id).eval().to(device)

waveform, sr = torchaudio.load("sample.wav")
if sr != 16000:
    waveform = torchaudio.functional.resample(waveform, sr, 16000)

inputs = processor.feature_extractor(waveform.mean(dim=0).numpy(), sampling_rate=16000, return_tensors="pt")
input_features = inputs.input_features.to(device=device, dtype=dtype)

predicted_ids = model.generate(input_features, language="Korean", task="transcribe", temperature=0.0)
print(processor.batch_decode(predicted_ids, skip_special_tokens=True)[0])
Read moreDetails
Benchmarks: 1.17% overall CER (down from 2.90% on the base model), and 0.59% on dysarthric speech (down from 4.45%).
Hardware: trained for 10 hours on 1x NVIDIA A100 (80GB VRAM).
Architecture: PEFT LoRA (r=64, lora_alpha=128), 16 kHz mono audio up to 29.5 seconds.
Limitations: not intended for clinical diagnosis.
Discuss Telegram 0 18

Will it run?

Runs in F16

Fits without quantisation.

  • F16weights 1.7 GB · ready files2.4 GB
  • FP8 / INT8recommendedweights 1.0 GB1.6 GB
  • 4-bitweights 0.6 GB1.2 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: FP8 / INT8 — from RTX 2060 · F16 — from RTX 2060

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

This is a LoRA: counted together with its base model whisper-large-v3-turbo.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # PL-HFT/whisper-large-v3-turbo-ko-dysarthria-lora
    # ===== Tokens: paste yours between the quotes =====
    export HF_TOKEN=""        # huggingface.co/settings/tokens, for gated models
    # ==========================================
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download openai/whisper-large-v3-turbo --local-dir ~/models/whisper-large-v3-turbo
    hf download PL-HFT/whisper-large-v3-turbo-ko-dysarthria-lora --local-dir ~/models/whisper-large-v3-turbo-ko-dysarthria-lora
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to whisper-large-v3-turbo
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

Requires the base model

openai/whisper-large-v3-turbo

This model is a LoRA / adapter. Download the base model first, then load this one on top.

  1. whisper-large-v3family root
  2. whisper-large-v3-turbo
  3. whisper-large-v3-turbo-ko-dysarthria-lora

New posts about versions of whisper-large-v3 (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 3All on Hugging Face

About the model

PL-HFT/whisper-large-v3-turbo-ko-dysarthria-lora
Author
PL-HFT
Category
Audio
Released
Posted

On NeyroLearn

Hugging Face

180

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments