Skip to content
Audio

AksharSTT v1

офлайн-STT для хинди и Hinglish на CPU

2 min readupdated

AksharSTT v1 от byteknotlabs — распознавание речи для хинди, английского и Hinglish, рассчитанное на обычные CPU и телефоны. Версия 1 — один файл на 29,5 МБ с 4-битными весами, работает через движок Cactus needle. База — Cactus Whistle, учитель — Shruti.

Ключевые особенности

  • Работает офлайн на обычных CPU и телефонах.
  • Один файл 29,5 МБ, 4-битные веса.
  • Хинди, английский и Hinglish; для хинди и Hinglish — тег hi, для английского — en.
  • Архитектура Cactus Whistle: 8 encoder и 8 decoder слоёв шириной 512, около 55 млн параметров; дистилляция из Shruti.
  • Один стиль вывода: хинди деванагари, английские слова строчной латиницей, числа словами, без пунктуации; пунктуацию и цифры оставляет приложению.
  • transcribe принимает WAV или 16 кГц mono float samples в [-1, 1] любой длины; с word_timestamps=True возвращает старт, конец и вероятность каждого слова.

Пример

pip install cactus-needle==3.2.8 huggingface_hub

from huggingface_hub import hf_hub_download
from needle import Whistle

model = Whistle(hf_hub_download('byteknotlabs/AksharSTT-V1', 'akshar-v1.cact'))
print(model.transcribe('speech.wav', language='hi')['text'])
ПодробнееСредний WER по шести тестам — 22,28% против 21,78% у Shruti; лучше на FLEURS Hindi (19,37 против 20,21) и MUCS Hinglish (34,41 против 35,26), хуже на Svarah Indian English (24,68 против 22,45).
Обучение — около 11,5 часов на одной RTX 5080; около 55 млн параметров; лицензия Apache-2.0.
Обучена и протестирована только на хинди и английском; другие языковые теги Whistle остаются в словаре, но не тестировались. Оценки сняты на engine build 3.3.3.
Discuss Telegram 1

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 0.2 GB · ready files0.7 GB
  • FP8 / INT8recommendedweights 0.1 GB0.6 GB
  • 4-bitweights 0.1 GB0.6 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 · BF16 — 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.

    # byteknotlabs/AksharSTT-V1
    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 byteknotlabs/AksharSTT-V1 --local-dir ~/models/AksharSTT-V1
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to AksharSTT-V1
  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 whistle (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 5All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 2All on Hugging Face

About the model

byteknotlabs/AksharSTT-V1
Author
byteknotlabs
Category
Audio
Released
Posted

On NeyroLearn

Hugging Face

01

Ask AI about this model

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

Comments

Audio

Swanada

Swanada