Skip to content
Audio

Whisper large-v3 copy-heads-off

меньше зацикливаний без обучения

2 min readupdated

Это Whisper large-v3 (1.54B) с отключёнными 11 головами decoder self-attention, которые копируют предыдущий текст и вызывают повторы на длинных записях. Ничего не дообучали: веса изменены одной арифметической правкой, остальные тензоры, имена и fp16 совпадают с оригиналом. Ориентирована на длинные аудио на датском и английском, особенно через whisper.cpp.

Ключевые особенности

  • Обнулены 64 столбца out_proj у 11 голов decoder self-attention; изменены 9 из 1259 тензоров, без тренировки.
  • На 30 скрытых датских записях длинные повторы сократились с 21 до 1, слова из повторов — на 94%.
  • Случайные правки такого же размера не дали похожего эффекта; FLEURS da_dk WER 12.56% → 12.45%.
  • Строгий порог почти достигнут: нужно −80% повторов, получено 76%; Lost Audio 15 против 14.
  • Есть fp16 safetensors и ggml f16/q5_0; на q5_0 число повторов на held-out упало на 79%.
  • Не лечит выдуманный текст в тишине — нужен VAD; на коротких аудио выигрыша нет.

Пример

import torch
from transformers import pipeline

asr = pipeline("automatic-speech-recognition",
    model="emillykkegrann/whisper-large-v3-copy-heads-off",
    torch_dtype=torch.float16, device="cuda")
out = asr("long_recording.wav", return_timestamps=True)
print(out["text"])
ПодробнееПравка: обнулены 64 столбца out_proj у 11 голов decoder self-attention; изменены 9 из 1259 тензоров, обучения не было.
На длинных датских записях WER 31.0% → 19.7% (timestamps), на английских длинных клипах Codictate 17.6% → 1.7%.
Тесты в основном whisper.cpp/crispasr на Apple Silicon; английский проверен на 8 длинных клипах.
Ограничения: нужен VAD, на коротких аудио выигрыша нет, датский WER не улучшается.
Discuss Telegram 0 55

Will it run?

Runs in F16

Fits without quantisation.

  • F16weights 2.9 GB · ready files3.7 GB
  • FP8 / INT8recommendedweights 1.4 GB2.1 GB
  • 4-bitweights 0.8 GB1.4 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.

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.

    # emillykkegrann/whisper-large-v3-copy-heads-off
    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 emillykkegrann/whisper-large-v3-copy-heads-off --local-dir ~/models/whisper-large-v3-copy-heads-off
    # The model is in ~/models. How to run it is in the model card (link above).
  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

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

emillykkegrann/whisper-large-v3-copy-heads-off
Author
emillykkegrann
Category
Audio
Released
Posted

On NeyroLearn

Hugging Face

550

Ask AI about this model

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

Comments