EmbeddingGemma 2 fine-tuned for 3 hours tops 8 Vietnamese ASR benchmarks at 356M params
stablequan · x · 2026-10-23
Google's Gemma team amplified a community result: fine-tuning EmbeddingGemma 2 on just 3,000 hours of Vietnamese audio for 3 hours tops 8 Vietnamese ASR benchmarks, without a heavy autoregressive decoder.
Benchmark charts show the 356M-param EmbeddingGemma2-ASR-vi averaging 6.87% WER, beating Gemini-Flash-3.8 (6.92%), MAI-Transcribe-2.0, Cohere Transcribe, Qwen3-ASR-1.7B and the entire Whisper family, with a low of 1.62% WER on bud500.
More from Models
- 400 LoRA steps on Gemma 4 31B beat dedicated decision model Jev — vanstriendaniel · 2026-10-24
- Kimi K3.1 Pro goes live on Kimi Code at a reported 117 tok/s — kimmonismus · 2026-10-24
- Community note debunks claim that distant tokens never share MoE experts — NathanpmYoung · 2026-10-24
- FreshStack retrieval leaderboard revamped with Cohere Embed v5 and Perplexity models — beirmug · 2026-10-24
- Kimi K3.1 Pro flagship model spotted rolling out on Kimi web and mobile — testingcatalog · 2026-10-24
- Qwen3.8-27B at 1.8bpw collapses to 13% on MiMo RL bench, EXL3 tests show — casper_hansen_ · 2026-10-24