On-Device AI — Apple Neural Engine
Two neural networks compiled to CoreML. Running on Apple Neural Engine. No cloud, no API keys, no accounts. Speech recognition that rivals cloud services — entirely on your device.
NVIDIA Parakeet v3 for English and European languages. Qwen3 ASR for 30+ languages. Both open-weight, both running locally, both discarding audio the moment transcription completes.
< 4% WER
English accuracy
Parakeet v3 on LibriSpeech test-clean
< 100ms
First-token latency
On Apple M1 Pro with ANE acceleration
0.08–0.15x
Real-time factor
Processes faster than real-time speech
200–700 MB
Memory footprint
Depending on selected model
Architecture
Every Apple Silicon Mac has a dedicated Neural Engine — a hardware accelerator designed for machine learning inference. Resonant compiles speech models to CoreML format, which runs on the Neural Engine directly.
This means transcription doesn't compete with your GPU for resources. You can run Resonant while gaming, rendering video, or training models — the Neural Engine operates independently.
Audio is captured, transcribed, and discarded in a single pass. No audio file is written to disk. No buffer is retained after processing. The raw waveform exists only in memory for the duration of transcription.
Audio lifecycle
Capture
PCM audio from microphone — held in memory buffer
Process
CoreML model runs on Neural Engine — tokens emitted
Format
19-stage Rust pipeline — filler removal, ITN, punctuation
Discard
Audio buffer released — no file written, no data retained
Pipeline
Six stages, all on your Mac. Audio capture through paste in under a second.
01
Audio capture
Raw PCM from mic or system audio
02
VAD
Silero VAD — detects speech boundaries
03
Neural Engine
CoreML model runs on ANE
04
Raw transcript
Token sequence with timestamps
05
Text pipeline
19-stage Rust formatter
06
Paste
Clean text in active field
Models
No black-box APIs. Every model Resonant ships is open-weight, published, and benchmarked. You know exactly what's processing your voice.
Both models are bundled with Resonant — no download on first launch. See all supported languages.
Comparison
Trust model
Resonant has no backend infrastructure for speech processing. There's no server to receive your audio, no database to store transcripts, and no pipeline to train on your data. The privacy claim isn't a policy — it's an architectural constraint.
You can verify this yourself. Run any network monitor while using Resonant. Dictate anything. No outbound connections during transcription. The only network traffic is optional — checking for updates and cloud text cleanup, both of which you can disable.
For regulated environments, Resonant includes HIPAA mode — encrypted local storage, suppressed telemetry, and a full audit log of all processing events.
Free. Local. Verifiable.
Two neural networks on your Mac. No cloud. No data retention. No subscription. Just local AI that works.
Requires macOS 14+ · Apple Silicon