Skip to content
Audio

gemma-stt-hindi

компактная ASR для хинди со скоростью 140× реального времени

2 min read

Небольшая модель распознавания речи для хинди от автора bringagi: аудиоэнкодер EmbeddingGemma 2 (12-слойный conformer, 305M параметров), у которого дообучены только верхние 2 слоя, плюс 6-слойный CTC-«ридер» (d=768), выдающий символы деванагари. Обучена дистилляцией двух больших моделей хинди-ASR на бесплатных Kaggle T4. Авторегрессивного декодера нет — весь клип транскрибируется за один forward pass.

Ключевые особенности

  • ~40 мс на клип 5.6 с на RTX 3050 Laptop (4 ГБ), это примерно 140× быстрее реального времени; 1.2 ГБ VRAM в fp16
  • Опциональный KenLM beam search (3-gram) добавляет ~10 мс CPU на клип и снижает число ошибок в словах примерно на 25%
  • 5.4% CER на Lahaja с LM — на уровне Qwen3-ASR-1.7B
  • Только хинди: на выходе деванагари без пунктуации и цифр, числа выводятся словами
  • Один forward pass вместо авторегрессии — быстрый инференс даже на слабом железе

Пример

from huggingface_hub import snapshot_download
import sys; sys.path.insert(0, snapshot_download("bringagi/gemma-stt-hindi"))
from gemma_stt import GemmaSTT

stt = GemmaSTT.from_pretrained("bringagi/gemma-stt-hindi")
print(stt.transcribe("clip.wav"))          # greedy CTC
print(stt.transcribe("clip.wav", lm=True)) # + 3-gram LM
ПодробнееБенчмарки: 5.4% CER / 15.9% WER на Lahaja-100 с LM (greedy — 6.4% / 21.3%), 4.87% CER на Kathbath-valid.
Обучение: ~1750 ч транскриптов от учителей, 605 минут на одной Kaggle T4 в fp16.
Ограничения: слабее на разговорной речи с ноутбучного микрофона (25% CER), не тестировалась на шумном, дальнем и телефонном аудио.
Лицензия: Indic Open Model License v1.0 — коммерческое использование и self-hosting разрешены, для публичного API нужно письменное одобрение Bodhan AI.
Discuss Telegram 0 26

Will it run?

Runs in F16

Fits without quantisation.

  • F16weights 0.6 GB · ready files1.2 GB
  • FP8 / INT8recommendedweights 0.3 GB0.9 GB
  • 4-bitweights 0.2 GB0.7 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.

    # bringagi/gemma-stt-hindi
    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 bringagi/gemma-stt-hindi --local-dir ~/models/gemma-stt-hindi
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to gemma-stt-hindi
  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 embeddinggemma-2 (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face

About the model

bringagi/gemma-stt-hindi
Author
bringagi
Category
Audio
Released
Posted

On NeyroLearn

Hugging Face

260

Ask AI about this model

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

Comments