---
license: apache-2.0
base_model: Qwen/Qwen3-1.7B
tags:
  - gguf
  - speech-to-text
  - text-cleanup
  - dictation
  - lora
  - qwen3
language:
  - en
pipeline_tag: text-generation
---

# TurboSpeak Correction Model

Fine-tuned Qwen3-1.7B (Q4_K_M quantization) for cleaning up speech transcription output.

## What it does

- Removes filler words (um, uh, like, you know, basically)
- Fixes stutters (w-w-want → want)
- Resolves mid-sentence self-corrections (speaker says X then corrects to Y → keeps only Y)
- Preserves all content words — never adds words the speaker didn't say

## Performance

| Metric | Score |
|--------|-------|
| Correction accuracy | 87.5% (35/40 P+G) |
| Filler/stutter handling | 100% |
| Avg latency | ~100ms on Apple Silicon |
| Model size | 1.0 GB (Q4_K_M) |

## Training

- Base model: Qwen/Qwen3-1.7B
- Fine-tuning: LoRA (rank=8, lr=5e-5, 500 iterations)
- Training data: 2,390 examples (1,710 base + 680 hard corrections)
- Quantization: Q4_K_M via llama.cpp

## Usage

Used by [TurboSpeak](https://github.com/kavanaghpatrick/gpu-forge) macOS dictation app.
Runs locally via llama.cpp / llama-cpp-2 Rust bindings.

### System prompt (ChatML format)

```
Clean up the transcribed text. Remove filler words, fix stutters, and resolve mid-sentence corrections. Output only the cleaned text.
```

## License

Apache 2.0 (same as base Qwen3 model)