Skip to content
Audio

w2v-BERT 2.0 для киньяруанда

CTC-ASR для нестандартной речи

2 min read

Это CTC-модель для распознавания речи на киньяруанда (rw), дообученная на одну эпоху от badrex/w2v-bert-2.0-kinyarwanda-asr. База — w2v-BERT 2.0 с CTC-головой, около 600 млн параметров; дообучение шло на CDLI non-standard speech dataset с преимущественно молодыми взрослыми, которые заикаются. Автор предупреждает: это ранняя версия, лучше использовать oluwagbotty/w2vbert-ky-ft-3ep.

Ключевые особенности

  • CTC-архитектура: не застревает в циклах повторения.
  • Вывод — строчные буквы киньяруанда, пробелы и апострофы, без пунктуации.
  • На тесте CDLI: WER 42.8%, CER 12.8%; до дообучения — WER 49.7%, CER 15.2%.
  • Обучение всего 1 эпоха (721 шаг), автор советует 3-эпохную версию.
  • Выход дословный: повторы заикания часто сохраняются в тексте.

Пример

import soundfile as sf, torch
from transformers import AutoProcessor, Wav2Vec2BertForCTC
repo = "oluwagbotty/w2vbert-ky-ft"
p = AutoProcessor.from_pretrained(repo); m = Wav2Vec2BertForCTC.from_pretrained(repo).eval()
a, r = sf.read("clip_16khz_mono.wav")  # 16 kHz, mono
i = p(a, sampling_rate=16000, return_tensors="pt")
with torch.no_grad(): l = m(**i).logits
print(p.batch_decode(l.argmax(-1))[0])
ПодробнееТест: 1 169 записей, 9 спикеров; WER 42.8%, CER 12.8%, per-utterance WER 44.6%.
Валидация: WER 44.05% на шаге 500 (эпоха 0.69), 45.3% после полной эпохи.
Ограничения: всего 9 тестовых спикеров, различия менее ~3 WER пунктов не подтвердить; около двух третей ошибок — варианты орфографии и пробелов, поэтому CER честнее.
Данные: train — 11 536 записей, 46 спикеров, 47.7 часа; модель не для диагностики речи и решений о людях.
Discuss Telegram 0 28

Will it run?

Runs in F32

Fits without quantisation.

  • F32weights 2.2 GB · ready files2.9 GB
  • BF16weights 1.1 GB1.7 GB
  • FP8 / INT8recommendedweights 0.5 GB1.1 GB
  • 4-bitweights 0.3 GB0.8 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 · F32 — 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.

    # oluwagbotty/w2vbert-ky-ft
    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 oluwagbotty/w2vbert-ky-ft --local-dir ~/models/w2vbert-ky-ft
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to w2vbert-ky-ft
  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

  1. w2v-bert-2.0family root
  2. w2v-bert-2.0-kinyarwanda-asr
  3. w2vbert-ky-ft

New posts about versions of w2v-bert-2.0 (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 1All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 1All on Hugging Face

About the model

oluwagbotty/w2vbert-ky-ft
Author
oluwagbotty
Category
Audio
Released
Posted

On NeyroLearn

Hugging Face

280

Ask AI about this model

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

Comments