Grammar Correction: Local LLM Polish After Transcription
Optional on-device grammar and punctuation cleanup via Llama-3.2-3B. Faster than re-reading, fully offline, no cloud LLM cost.
Raw transcription captures what you said — but spoken language is messier than written text. SuperSpeech includes an optional grammar correction pass powered by a locally-running Llama-3.2-3B model. Filler words get removed, punctuation is tidied, sentences become complete. All on your device, adding only ~275ms to the pipeline.
Why post-processing matters
Raw ASR output often reads as a stream of consciousness. "um we should probably like go with option a because uh it is cheaper." Post-processing normalizes this: "We should go with option A because it is cheaper." For professional writing, that is a 2-3x productivity win over manual cleanup.
- Filler word removal: um, uh, like, you know
- Automatic punctuation and capitalization
- Sentence completion for fragment thoughts
- Preserves your wording — does not rewrite for style
- Configurable aggressiveness: light, balanced, or strict
How it works
After the ASR model (Parakeet) produces raw text, the output is optionally passed to a locally-hosted Llama-3.2-3B-Instruct model quantized to Q4 (~1.5GB download). The LLM runs via MLX on Mac (Apple Silicon optimized) or LLamaSharp on Windows (CUDA/DirectML/CPU). Latency is typically 200-300ms.
Supported languages
Grammar correction works well in English, German, French, Spanish, Italian, and Portuguese — the languages officially supported by Llama 3.2. Other languages either pass through unchanged or get minimal correction.
Privacy preserved
Like ASR, grammar correction runs locally. Your raw and corrected text never leaves the device. This is a meaningful difference from tools that send transcripts to OpenAI, Anthropic, or Google for cleanup — your confidential dictation stays confidential.
Frequently Asked Questions
Is grammar correction mandatory?
No. It is an optional post-processing step. You can toggle it per session or disable it entirely in settings.
Does it rewrite my content?
No. It corrects grammar, punctuation, and filler words while preserving your word choices. It does not paraphrase or change tone.
How fast is it?
On Apple Silicon M1-M4: typically 200-300ms added to transcription. On Windows with CUDA: similar. On CPU-only: 500-800ms.
Which languages are supported?
English, German, French, Spanish, Italian, and Portuguese have full support. Other languages pass through with minimal changes.
Do I need to download another model?
Yes. Grammar correction requires the ~1.5GB Llama-3.2-3B model. It is optional — skip it if you do not want the post-processing feature.
Clean transcripts on autopilot
Grammar correction is included in every plan. Enable in settings.