Skip to content
Audio

Parakeet TDT 0.6B v3

турецкий экспериментальный LoRA-адаптер

2 min read

ArdaOkan адаптировал NVIDIA Parakeet TDT 0.6B v3 для распознавания турецкой речи. В основе — rank-8 LoRA по вниманию во всех 24 слоях энкодера, при этом декодер, проекторы энкодер/декодер и joint-голова обучены полностью. Релиз экспериментальный: выбран чекпоинт update 400 000 с лучшим Turkish DEV WER.

Ключевые особенности

  • 192 LoRA-тензора rank-8 (alpha 16, dropout 0) плюс 15 полностью обученных тензоров декодера, проекторов и joint-головы.
  • Приоритет у турецкого, с реплеем английского, испанского, итальянского, французского, немецкого и португальского; остальные языки базовой модели не оценивались.
  • Добавлены турецкие символы и BPE-куски из корпуса, токенизатор расширен до 9 222 лексических единиц.
  • Веса лежат в adapter_model.safetensors (~84,1 MB); это нативный формат NeMo, а не PEFT-адаптер.
  • Лицензия CC BY-NC 4.0: только некоммерческое использование, с атрибуцией и указанием изменений.

Пример

from huggingface_hub import snapshot_download
import importlib.util, torch
folder = snapshot_download("ArdaOkan/parakeet-tdt-0.6b-v3-tr-experimental-lora")
spec = importlib.util.spec_from_file_location("parakeet_adapter", folder + "/load_adapter.py")
loader = importlib.util.module_from_spec(spec)
spec.loader.exec_module(loader)
model = loader.load_model(folder, device="cpu", cache_dir="./model_cache")
print(model.transcribe(["audio.wav"], batch_size=1)[0].text)
ПодробнееFLEURS TEST: турецкий WER упал со 100,96% до 18,21%; шесть языков реплея — с 4,59% до 5,01% (+0,42 pp).
Нативная merged-версия: Turkish DEV WER 22,06% (182/825 слов); 70/70 smoke-транскриптов совпали с merged-выходом.
Валидация шла на CPU FP32; GPU-попытки упёрлись в 8 GB host-memory guard и не считаются завершёнными.
Только исследовательское некоммерческое использование: возможны пропуски, замены и выдуманные слова, особенно при domain shift.
Discuss Telegram 0 112

Will it run?

Runs in F32

Fits without quantisation.

  • F32weights 2.4 GB · ready files3.2 GB
  • BF16weights 1.2 GB1.9 GB
  • GGUF Q8_0recommendedweights 0.7 GB · ready files1.3 GB
  • FP8 / INT8weights 0.7 GB1.2 GB
  • 4-bitweights 0.4 GB0.9 GB

Tip: go 8-bit.GGUF Q8_0 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: GGUF Q8_0 — from RTX 2060 · F32 — 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 parakeet-tdt-0.6b-v3.

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.

    # ArdaOkan/parakeet-tdt-0.6b-v3-tr-experimental-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 nvidia/parakeet-tdt-0.6b-v3 --local-dir ~/models/parakeet-tdt-0.6b-v3
    hf download ArdaOkan/parakeet-tdt-0.6b-v3-tr-experimental-lora --local-dir ~/models/parakeet-tdt-0.6b-v3-tr-experimental-lora
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to parakeet-tdt-0.6b-v3
  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

nvidia/parakeet-tdt-0.6b-v3

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

New posts about versions of parakeet-tdt-0.6b-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 1All on Hugging Face

About the model

ArdaOkan/parakeet-tdt-0.6b-v3-tr-experimental-lora
Author
ArdaOkan
Category
Audio
Released
Posted

On NeyroLearn

Hugging Face

1120

Ask AI about this model

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

Comments