Vikhrmodels/salt-qwen2.5-0.5b-asr

Model Performance Overview

Metrics:

• CER: Character Error Rate (lower = better).
• WER: Word Error Rate (lower = better).

Model	CER	WER
SALT-asr	8.42	18.49

---

Our Solution

• Method: Extends a pre-trained LLM with audio tokens and fine-tunes on ASR task.
• Audio tokenization: SpeechTokenizer (semantic tokens only).

---

Resources

• Code: GitHub Repo

---