Speech Models

VoxFlow supports Apple Speech out of the box, local ASR providers for offline workflows, and cloud providers for users who prefer hosted recognition. The app labels local versus cloud, streaming support, language coverage, and readiness in the Models UI.

Local And System Providers

Local providers keep audio on the Mac, except for Apple Speech behavior governed by macOS system services.

Provider / Model	Status	Streaming	Runtime Route	Recommended Use
Apple Speech	Supported	Yes	Apple Speech / SFSpeechRecognizer	Start immediately without downloading a model
Qwen3-ASR 0.6B	Supported	Partial + final	speech-swift Qwen3ASR / MLX 4bit	Default local dictation route with balanced size and speed
Qwen3-ASR 1.7B	Supported	Partial + final	speech-swift Qwen3ASR / MLX 8bit	Higher-accuracy local dictation when memory allows
Whisper Turbo / Large V3	Supported	No	WhisperKit .mlmodelc	High-quality full-recording transcription after capture ends
FunASR	Supported	Segment confirmation	Sherpa-ONNX	Local Chinese fallback route, not CoreML
SenseVoice	Supported	Short utterance / non-streaming path	FluidAudio / CoreML	Local multilingual short-utterance transcription
Paraformer Large zh	Supported	Segment confirmation	FluidAudio / CoreML int8	Local Chinese transcription
NVIDIA Nemotron 0.6B	Supported	Yes	speech-swift NemotronStreamingASR / CoreML	Local multilingual streaming candidate
Parakeet Streaming	Supported	Yes	speech-swift ParakeetStreamingASR / CoreML	Low-latency English and European-language dictation
Omnilingual ASR	Supported	No	speech-swift OmnilingualASR / CoreML	Broad-language offline transcription and experimental workflows

Cloud Providers

Cloud providers send recorded audio to the selected service. Credentials are stored in the local credentials file.

Cloud Provider	Status	Streaming	Default Model / API	Configuration
Groq	Supported	No	whisper-large-v3-turbo audio transcription	API key and model
Tencent Cloud	Supported	Yes	Realtime Speech Recognition WebSocket, 16k_zh	AppID, SecretId, SecretKey
Alibaba Cloud	Supported	Yes	DashScope WebSocket, fun-asr-realtime	Bailian API key
Volcengine Cloud	Planned	Planned	Doubao streaming ASR	To be determined
Mistral Voxtral	Not yet supported	To be determined	Voxtral speech capability	None
AssemblyAI	Not yet supported	To be determined	AssemblyAI Transcription	None
ElevenLabs Scribe	Not yet supported	To be determined	ElevenLabs Scribe	None

Selection Semantics

VoxFlow only persists a selected provider after it is selectable. If a local model is missing or a cloud credential is not configured, selecting that provider is rejected instead of silently persisting a provider that will fall back at runtime.

Older persisted settings can still fall back to Apple Speech when the selected provider becomes unavailable, such as after deleting a local model directory.

Choosing A Provider

• Start with Apple Speech if you want the quickest setup.
• Use Qwen3-ASR 0.6B for local everyday dictation.
• Use Qwen3-ASR 1.7B when accuracy matters more than model size.
• Use Whisper for high-quality file or recording transcription where real-time feedback is less important.
• Use Tencent Cloud or Alibaba Cloud when you need cloud streaming.
• Use Groq when you want hosted final transcription without local model downloads.