There is no single best speech model
A quiet English call, a multilingual interview, and a noisy recording with domain vocabulary are different problems. LokalBot keeps several engines available so you can choose a fast default and retain a broader or more accurate fallback. All of the options below run locally after their files have downloaded.
| Model | Best fit | Coverage / size |
|---|---|---|
| IBM Granite Speech 4.1 2B | Alternative local speech engine | Local llama.cpp speech model |
| Parakeet TDT 0.6B v3 | Very fast multilingual meetings | 25 languages; ~190× realtime in project benchmarks |
| Parakeet TDT 0.6B v2 | English-focused recall | English only |
| Qwen3-ASR 1.7B | Default in LokalBot 0.8.1 | 52 languages/dialects; ~3.2 GB |
| Qwen3-ASR 0.6B | Compact broad coverage | Global coverage; ~0.7 GB |
| Whisper large-v3 turbo | Wide-language fallback and timestamps | 99 languages; ~1.6 GB |
A sensible selection strategy
- Start with Qwen3-ASR 1.7B for the default speech model in LokalBot 0.8.1.
- Choose Parakeet v3 when throughput and its 25 supported languages cover your meetings.
- Keep Whisper installed if you need broader language coverage or word timestamps.
- Compare Granite Speech as another local option on the recordings that matter to you.
- Compare on your audio. A two-minute representative clip is more informative than a generic benchmark.
Speech recognition is only one stage
Speaker separation, punctuation, summary quality, and action-item extraction depend on later stages too. LokalBot begins with separate microphone and selected-app system-audio tracks, whose labels indicate capture source rather than speaker identity. It can apply on-device diarization to system audio before sending the transcript to the selected summarization backend. A perfect language model cannot recover words that the speech model never recognized, so improve capture and transcription before tuning recap prompts.
Storage and memory planning
Speech models are not the only downloads. Local summary and cotyping models range from about 0.53 GB to roughly 17.73 GB in the built-in catalog. Smaller options work on any supported Apple Silicon Mac; several quality-focused choices recommend 16 GB, while the largest long-meeting defaults target 32 GB or more. Install only what you use and keep free space for recordings.
How to evaluate output
- Use the same audio file for every candidate.
- Check names, numbers, domain terms, and language switches.
- Review timestamps and speaker boundaries, not just prose readability.
- Measure end-to-end time on your own Mac.
- Keep the model whose errors are easiest for your workflow to notice and fix.