$78 fine-tune of EmbeddingGemma 2 beats Gemini at Vietnamese speech recognition
aigclink · x · 2026-10-15
A community project repurposed Google's EmbeddingGemma 2 audio encoder into a SOTA Vietnamese ASR model for just $78: fine-tuning only the top 2 of 12 encoder layers and training a 6-layer Transformer reader with CTC on 3,000 hours of Vietnamese audio, using 8×H100 for about 3 hours.
- Average WER of 6.87% across 8 Vietnamese benchmarks, edging out Gemini-Flash-3.8 (6.92%), Cohere Transcribe (7.01%), and ElevenLabs Scribe v2 (7.08%)
- No autoregressive decoding: 200× real-time on CPU, supports Vietnamese-English code-switching and streaming (0.48s chunks)
- The same embedding space earlier powered a macOS menu-bar local semantic search tool, mapping text/image/audio/video into one 768-dim space via llama.cpp+Metal with hybrid vector+keyword retrieval
If this recipe generalizes across languages, it could reshape niche speech-recognition markets for underrepresented languages and dialects.
Related event: $78 and 2 hours on 8 H100s yields Vietnamese ASR SOTA(4 posts)→
More from Multimodal
- Midjourney v8.2 style test: bold flat-ink illustration looks, sref code included — azed_ai · 2026-10-15
- Sarvam AI launches Bulbul V4 Flash, a faster compact TTS model for Indian languages — CurieuxExplorer · 2026-10-15
- Maya-2 debuts as open-weight Indic voice model, ranks #2 on Hindi TTS leaderboard — CurieuxExplorer · 2026-10-15
- PixText fixes text in generated images: 83% glyph pass rate vs 37% for baselines — ATH-MaaS · 2026-10-15
- TextVerifier catches flawed Chinese glyph rendering with F1 0.844, far above OCR and VLMs — ATH-MaaS · 2026-10-15
- AI Video Creators: 'Amazing Visuals Are Solved — Unique IP and Characters Are What's Left' — AIandDesign · 2026-10-15