beingpax/voiceink-refine-v1

VoiceInk Refine v1: ASR Transcript Cleanup Model

VoiceInk Refine v1 is a specialized 5.1 billion parameter language model, fine-tuned by beingpax from google/gemma-4-E2B-it. Its primary function is to refine noisy Automatic Speech Recognition (ASR) transcripts, converting them into clean, readable text without altering the speaker's original meaning.

Key Capabilities

• ASR Transcript Refinement: Specifically designed to clean up raw ASR outputs.
• Meaning Preservation: Focuses on improving text quality while maintaining the semantic integrity of the original speech.
• Efficient Training: Trained using 16-bit LoRA with 81,786 examples over one epoch, achieving a final average training loss of 0.08115.
• Optimized for Latency: Inference can be performed with enable_thinking=False for low-latency cleanup.

Training Details

The model was fine-tuned from google/gemma-4-E2B-it using 16-bit LoRA. Key training parameters included a maximum sequence length of 2,048, a learning rate of 2e-4 with cosine scheduling, and an assistant-response-only loss function. Evaluation metrics such as WER, CER, exact-match, and overcorrection are still pending.

When to Use This Model

This model is ideal for applications requiring high-quality text output from ASR systems. It is particularly suited for:

• Post-processing ASR transcripts: Improving the readability and accuracy of automatically generated text.
• Specialized ASR cleanup: For scenarios where preserving the original meaning is critical.

Users should evaluate the model's performance on their specific languages, accents, punctuation conventions, and proper nouns before deploying it in production, as it is specialized for VoiceInk-style ASR cleanup.