Audio8-ASR-Infinite-80ms for Agora

This public repository contains the 80ms streaming ASR package. Application code and deployment guide are maintained separately. Upstream model: Edge0/Audio8-ASR-Infinite.

Resource layout on main

models.json / agora-model.json / SHA256SUMS
embedding.bin / *.npy / tokenizer.json / token_bytes.json
frontend/ prefix/ semantic_vad/ LICENSES/ NOTICE
targets/
  QCS8550/
  QCS8625/
  IQ8/
  IQ9/

Common resources are maintained once. Each SoC directory contains its hardware graphs. Branches are not used to select hardware. Model and code releases can be identified with version tags.

Download with the official Hugging Face CLI

Set only AGORA_WORKSPACE and AGORA_SOC directly in the device terminal. Choose any workspace path/name. No configuration file transfer is required:

export AGORA_WORKSPACE="$HOME/my-asr-demo"
export AGORA_SOC='QCS8550'

Follow the five-step deployment guide to clone code and prepare the environment before downloading resources.

Install huggingface_hub during dependency setup, and activate the project virtual environment once:

cd "$AGORA_WORKSPACE/code"
source .venv/bin/activate
pip install --upgrade huggingface_hub

Keep using this activated terminal. In a new terminal, set the same two variables, change to the code directory, and activate .venv again.

Download common resources

hf download APLUX-Agora/Audio8-ASR-Infinite-80ms \
  --exclude "targets/*" --local-dir "$AGORA_WORKSPACE/resources"

Download the SoC model

hf download APLUX-Agora/Audio8-ASR-Infinite-80ms \
  --include "targets/$AGORA_SOC/*" --local-dir "$AGORA_WORKSPACE/resources"

The default revision is main. Supported AGORA_SOC values: QCS8550, QCS8625, IQ8, IQ9. Public downloads require no token. Re-running the same commands reuses current files and downloads missing or changed files. No Git LFS installation, custom CLI, or sparse checkout is needed. The official HF CLI guide documents this download interface.

A selected package contains about 3.3 GiB of model data. Allow at least 8 GiB for the workspace, dependencies, and download overhead. Use a separate workspace per task or model/hardware combination. Existing unrelated files are not automatically removed.

Verification and use

cd "$AGORA_WORKSPACE/resources"
sha256sum -c SHA256SUMS --ignore-missing
cd "$AGORA_WORKSPACE/code"

Unselected SoC directories are intentionally absent. The checksum command validates downloaded files; missing-file tolerance alone does not prove completeness. The application's preflight checks required model files and hashes before startup. models.json remains the SDK entry point and owns application defaults. Only QCS8550 / Ubuntu 22.04 has been validated in the Agora lab; other hardware combinations need device qualification.

License

Original model, SDK, and third-party license terms are unchanged. See LICENSES/ and NOTICE. Full deployment instructions.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for APLUX-Agora/Audio8-ASR-Infinite-80ms

Finetuned
(4)
this model