Speech Recognition Models for Voxt
Voxt uses a separate ASR channel to turn speech or a recording into text. The right speech recognition model depends on language, audio quality, meeting features, model size, and whether you want the recognition step to stay on your Mac.
The short version
- Whisper is the general-purpose starting point for broad speech-to-text coverage.
- Qwen3 ASR, Parakeet, Nemotron, and SenseVoice provide additional local model choices for different languages and latency preferences.
- MOSS-Transcribe-Diarize is the specialist option for timestamped, multi-speaker meeting transcripts when the required workflow is available.
- Apple System Dictation is the lowest-friction option when you do not want to download a model.
- Availability, download size, language coverage, and output quality depend on the current Voxt build, selected channel, audio, and Mac hardware.
Model map
| Model or family | Role | Useful when | Upstream source |
|---|---|---|---|
| Apple System Dictation The no-download starting point for basic macOS dictation and quick voice input. | System ASR | Choose it when setup simplicity matters more than model control. | Apple Support |
| Whisper Large v3 Turbo, Large v3, and Small The main general-purpose Whisper family exposed through Voxt's local audio model path. | Local ASR | A sensible first comparison for multilingual dictation and recordings. | OpenAI Whisper |
| Qwen3 ASR 0.6B and 1.7B Qwen speech recognition variants available in multiple quantized sizes in the current model documentation. | Local ASR | Compare the smaller and larger variants when language coverage, model size, and recognition quality pull in different directions. | Qwen3-ASR |
| Parakeet v3 An NVIDIA NeMo speech recognition family represented in Voxt's MLX Audio model choices. | Local ASR | Worth testing for English-heavy workflows and users comparing local ASR families. | NVIDIA NeMo Speech |
| Nemotron 0.6B Streaming A streaming-oriented local ASR option in the current Voxt model documentation. | Streaming local ASR | Use it as a candidate when live preview and incremental recognition are important. | NVIDIA NeMo |
| MOSS-Transcribe-Diarize A model for long-form multi-speaker transcription with timestamps and speaker diarization. | Meeting ASR | A natural candidate for meeting recordings where who said what matters. | OpenMOSS |
| SenseVoice Small A multilingual speech understanding model with language identification and additional audio understanding capabilities upstream. | Multilingual ASR | Test it when multilingual recognition or language-aware audio workflows are more important than a single default model. | FunAudioLLM SenseVoice |
How to choose an ASR model
Start with the audio you actually produce. Short dictation, noisy meetings, bilingual product names, and long recordings can favor different models. Compare a small sample with the same microphone and keep notes on recognition errors, punctuation, latency, and memory use instead of relying on a generic model ranking.
ASR is separate from cleanup
Voxt can route speech recognition and language-model enhancement through separate channels. That means a fast local ASR model can produce the transcript while another local or remote model cleans punctuation, applies an App Branch rule, or turns the transcript into a finished message.
Meeting features need model and workflow support
A plain transcript does not automatically contain reliable speaker labels or timestamps. Treat diarization and meeting analysis as model- and workflow-dependent, and review names, decisions, and action items before sharing a meeting record.
FAQ
What speech recognition models does Voxt support?
The current Voxt model documentation includes Apple System Dictation, Whisper variants, Qwen3 ASR, Parakeet, Nemotron streaming ASR, MOSS-Transcribe-Diarize, and SenseVoice for local or system speech recognition. Remote ASR providers are documented separately and can depend on provider configuration and the current build.
Is Whisper the only local ASR option in Voxt?
No. Whisper is one general-purpose option. Voxt's documented local choices also include Qwen3 ASR, Parakeet, Nemotron, MOSS-Transcribe-Diarize, SenseVoice, and Apple System Dictation. Model visibility and download options can change with releases.
Which model is best for meeting transcription?
There is no universal best model without a controlled test. MOSS-Transcribe-Diarize is the most directly aligned with multi-speaker transcripts, timestamps, and diarization, while other ASR models may be a better fit for simple dictation or lower resource use.