Gemma 4 speech encoder matches larger models on arabic tasks with adapters
The researchers keep the encoder weights frozen and train a 22.6 M‑parameter CTC reader, then compare it to a version that adds 3.1 M low‑rank adapter parameters, a 41.5 M hybrid CTC/attention reader, and joint decoding with a sub‑word language model. Against Cohere Transcribe Arabic (2.07 B parameters) run on the same audio, the frozen system trails in word error rate but is statistically…
1 source primary source
Key points
- Frozen Gemma 4 encoder runs 3.4× faster on CPU than Cohere while matching CER on 5 of 6 Arabic conditions.
- Adapter‑augmented system reaches 19.20% WER on Saudi call‑center data, statistically indistinguishable from 19.27% baseline, and beats a 0.6B SADA model.
- Optimised joint decoder processes audio at 139× real‑time on one GPU, ranking 4th of 15 systems on a blind call‑center test.
Model page: EmbeddingGemma 2 →
Read the original at arXiv cs.CL · by Amr Abdelkhalik primary sourceOpen source ↗
Topics · follow one to build your own front page
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Research
All →- PHP patches critical RCE flaws discovered with GPT-5.6 Sol, affecting up to 70% of websites · 9 src
- OpenAI and Google AI models discovered critical PHP flaw affecting millions of websites · 8 src
- Researchers release madartb, a parallel Multi-Dialect Arabic dependency treebank · 1 src
- Researchers decode 26 legal factors in three open LLMs · 1 src
- Researchers test how reinforcement learning teaches AI agents to enforce market collusion · 1 src
Comments
via GitHub Discussions