What is Audio8-ASR-0.1B?

Audio8-ASR-0.1B is a compact automatic speech recognition model packaged for on-device iOS deployment. The ANE-oriented build uses Core ML for the audio tower and ONNX Runtime for an int4 language-model decoder.

Which languages does Audio8 ASR support?

The model supports English, Chinese, Cantonese, French, German, Japanese, and Korean in the published iOS ANE package.

When should I use on-device STT instead of a cloud API?

Choose on-device when privacy, offline use, or predictable latency matter more than the last point of accuracy on rare accents. Cloud APIs still win for 100+ language coverage and zero client hardware constraints.

Audio8 ASR 0.1B: 200MB on-device speech-to-text for seven languages

Cloud speech APIs are easy until you need offline intake, HIPAA-adjacent privacy, or sub-second latency on a bad hospital Wi-Fi network. That is where sub-billion-parameter ASR stacks are having a moment.

Audio8-ASR-0.1B showed up in the August 2026 digest as an open 0.1B speech-to-text model that runs fully on-device across seven languages. The iPhone ANE package on Hugging Face is the clearest public artifact: Core ML on Apple Neural Engine for the audio tower, ONNX Runtime for an int4 decoder, and a demo designed around ~200MB runtime memory on physical hardware.

I build voice agents for clinics and service businesses. This is the class of model I watch when a client says "we cannot send audio to OpenAI."

Why 0.1B ASR is a product decision, not a benchmark flex

Bigger is not always better for voice products. A front-desk agent that transcribes appointment requests on an iPad cares about:

Requirement	On-device 0.1B	Cloud STT
Privacy	Audio never leaves device	Audio hits vendor infra
Offline / flaky network	Works air-gapped	Fails or queues
Latency	ANE path can be very low	Network RTT added
Cost at scale	CapEx / device GPU	Per-minute OpEx
Language breadth	Narrow (7 here)	Often 100+
Accent robustness	Model-dependent	Often stronger at frontier

Audio8 targets the left column deliberately. Seven languages cover a large share of multilingual intake in North America and East Asia without shipping a 1B+ decoder to every phone.

On-device Audio8 ASR pipeline with Core ML audio tower and ONNX int4 decoder

What the iOS ANE package includes

From the Hugging Face model card:

Audio tower / head: compiled Core ML mlmodelc, mixed Float16/Int8 storage, ANE execution
Decoder: ONNX Runtime CPU decoder with int4 shared LM weights
Token embeddings: Float16 table on disk
Swift SDK + demo app for microphone transcription without network calls
Languages: English, Chinese, Cantonese, French, German, Japanese, Korean

Requirements skew Apple-native: macOS on Apple Silicon for builds, full Xcode, iOS 18+, physical device for realistic ANE numbers. The card shows example footprints around 183MB during live mic transcription, with peaks varying by device and cold start.

That memory budget matters. Many clinic iPads are not latest-gen pros. A 200MB-class stack leaves headroom for your UI, local RAG, and a small TTS model in the same process.

How this fits next to other local STT options

The ecosystem is crowded in a good way. Rough positioning:

Project	Size class	Strength
Audio8-ASR-0.1B (ANE)	~0.1B	iPhone ANE packaging, tiny RAM
IBM Granite 4.0 1B Speech	~1B	Enterprise multilingual ASR/AST
phonex (Sherpa-ONNX)	30-600MB models	Rust CLI, streaming, 10+ languages
OpenASR	Many families	Desktop OpenAI-compatible local API

Audio8 is not trying to be Whisper-large for 99 languages. It is trying to be the smallest usable iPhone transcription stack with a credible multilingual set.

For my voice ops work, the decision tree looks like:

Need 100+ languages or best accuracy? Cloud or Granite-class local.
Need streaming on server/Linux? phonex or vLLM-served models.
Need private on-iPhone capture? Audio8-class ANE packages.

Decision tree for choosing on-device versus cloud speech-to-text

Integration pattern for voice agents

A practical on-device voice loop for appointment booking:

VAD (voice activity detection) on device
Audio8 ASR for transcript
Small local LLM or rules engine for slot filling
Cloud escalation only when confidence is low or tools need CRM access

Keep PII-heavy audio local. Push structured fields (name, time window, service type) to your backend. That split satisfies a lot of clinic security reviews without giving up cloud CRM integrations.

Watch these failure modes:

Accent and domain vocabulary. Fine-tune or keyword-bias if drug names or trade terms matter.
Background noise. Front-desk environments are harsh; test with real HVAC and hold music bleed.
Language mixing. Cantonese/English code-switching is common; validate on real calls, not clean studio clips.

When I would not use Audio8

Skip on-device 0.1B ASR if:

You need real-time translation across many language pairs (look at AST models like Granite Speech)
Your users are mostly on Android or desktop browsers without an equivalent ANE package
Compliance requires vendor BAA and your cloud STT provider already offers one
Accuracy on rare dialects is the top metric and you have budget for cloud frontier models

Audio8 is a building block for privacy-first mobile capture, not a universal STT replacement.

Bottom line

Audio8-ASR-0.1B is part of a broader shift: speech stacks small enough to live beside your app, not behind a API key. For seven-language on-iPhone transcription at ~200MB, it is one of the most concrete open packages in the August digest.

If you are designing a voice product and debating cloud versus on-device, start with where audio is allowed to go, not which leaderboard model is biggest.

Building voice intake or on-device transcription into your ops stack? Book a free discovery call.