Transcription Models
Overview of local and cloud transcription models — speed, accuracy, and privacy trade-offs.
VivaDicta supports a wide range of transcription models — from fully local and private to lightning-fast cloud services.
Local Models
Local models run entirely on your Mac. No internet required, no data leaves your device.
Whisper
- OpenAI's Whisper model running locally on your Mac.
- Multiple model sizes: Tiny, Base, Small, Medium, Large, Large Turbo.
- Larger models are more accurate but require more RAM and processing time.
- Best on Apple Silicon Macs (M1 and later).
- Supports 100+ languages.
Parakeet
- NVIDIA's Parakeet model optimized for Apple Silicon.
- Supports 25+ languages with automatic language detection (v3).
- Fast and accurate — a great alternative to Whisper.
The Parakeet family gained three more members: Parakeet TDT-CTC-110M at 227 MB and roughly 100x realtime — the fastest local model available — plus Parakeet Unified, which reads better than V3 on English, and the two Nemotron builds (Latin and Multilingual), which trade about half the throughput for a lower word error rate. Nemotron requires Apple Silicon.
Apple Speech
- Uses macOS built-in speech recognition.
- No model download needed — works immediately.
- Good for quick dictation, less accurate for longer recordings.
Canary and Qwen3-ASR
Three more on-device models, run through the same local engine as Cohere's transcription model:
- Canary 180M Flash — 159 MB, English, German, Spanish and French. The instant tier: about a fifth of Whisper Turbo's download at comparable English accuracy.
- Canary 1B v2 — 735 MB, 25 European languages including Russian and Ukrainian.
- Qwen3-ASR 0.6B — 690 MB, auto-detect across 30 languages. The widest language coverage available locally.
- Cohere Local — 1.5 GB, 14 languages.
Cloud Models
Cloud models send your audio to a remote server for transcription. They're typically faster and more accurate than local models, especially on older Macs.
| Provider | Free Tier | Strengths | API Key |
|---|---|---|---|
| Groq | Free forever (rate-limited) | Fastest cloud transcription | Get key |
| Deepgram | $200 free credits | High accuracy (Nova-3) | Get key |
| ElevenLabs | Free tier (limited) | Great multi-language (Scribe 2) | Get key |
| Mistral (Voxtral) | Free Experiment plan | Open-source, multilingual | Get key |
| Google Gemini | Free tier + $300 credits | Google ecosystem | Get key |
| Soniox | Check their site | Real-time, diarization | Get key |
| OpenAI | Pay-as-you-go | Original Whisper API | Get key |
Newer Cloud Providers
- Gladia, Speechmatics and AssemblyAI — all three support speaker labels, and the first two can also translate while transcribing.
- Cartesia and xAI Speech-to-Text — xAI accepts a SuperGrok or X Premium sign-in in place of an API key.
- Gemini 3.5 Transcribe — 85+ languages, speaker labels, and it removes hesitation sounds and self-corrections as it transcribes. That clean-up can be switched off.
- Deepgram Flux — a turn-based realtime model that settles each turn as you finish speaking.
Every cloud model card carries badges showing whether it supports speaker labels and native translation, so choosing one for either job is not guesswork.
How to Choose
- Want free & fast? Groq — our #1 recommendation.
- Want maximum privacy? Local Whisper Large Turbo or Parakeet.
- Want best accuracy? Deepgram Nova-3 or ElevenLabs Scribe 2.
- On an older Mac? Cloud models offload compute — try Groq.
See Recommended Models for detailed recommendations by use case.