Skip to content
Audio

Zipformer-65M-vi-en

Bilingual Vietnamese and English speech recognition

2 min read

An offline automatic speech recognition (ASR) model for Vietnamese and English, trained from scratch without relying on existing pretrained checkpoints. Built on the Zipformer2 transducer (RNN-T) architecture using the icefall, k2, and lhotse frameworks, it stays compact at around 80.9M parameters. The model outputs properly capitalized text using a dedicated token in its BPE vocabulary. It is a solid fit for developers who need local recognition of mixed Vietnamese-English speech with inference support via sherpa-onnx.

Key features

  • Trained entirely from scratch (random initialization) in four stages: base training followed by three sequential fine-tuning runs.
  • Compact footprint of 80,931,511 parameters using a Zipformer2 transducer architecture without auxiliary CTC or attention heads.
  • 8,000-token SentencePiece BPE tokenizer includes a dedicated token for encoding uppercase letters.
  • Compatible with ONNX export and execution via sherpa-onnx (the repository includes the checkpoint, tokenizer model, and token table).

What you can do

  • Run local offline transcription for Vietnamese and English audio files using sherpa-onnx.
  • Use the model.pt checkpoint as a base model for further domain-specific fine-tuning in icefall.
  • Integrate it into voice services or assistants where users switch between Vietnamese and English mid-sentence.

Example

hf download ZAIC26/Zipformer-65M-vi-en model.pt bpe.model tokens.txt \
  --local-dir base_models/Zipformer-65M-vi-en
Read moreDetails
Input: 80-dimensional fbank features, 16 kHz mono audio.
Operating mode: offline only (non-streaming), without real-time streaming recognition.
The model.pt checkpoint was created by averaging the last 20 epochs of the final training stage (epochs 6–25).
Quality metrics (WER/CER) are not listed on the model card.
Discuss Telegram 0

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 0.3 GB · ready files0.8 GB
  • FP8 / INT8recommendedweights 0.2 GB0.7 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.

    # ZAIC26/Zipformer-65M-vi-en
    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 ZAIC26/Zipformer-65M-vi-en --local-dir ~/models/Zipformer-65M-vi-en
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to Zipformer-65M-vi-en
  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 Zipformer-65M-vi-en (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 1All on Hugging Face

About the model

ZAIC26/Zipformer-65M-vi-en
Author
ZAIC26
Category
Audio
Released
Posted

On NeyroLearn

Hugging Face

00

Ask AI about this model

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

Comments

Audio

AIAMD

AIAMD