QUAN-Transcribe-1.0: open Vietnamese ASR trained in 3 hours on 8×H100 beats Gemini
aigclink · x · 2026-10-15
QUAN-Transcribe-1.0, an open Vietnamese ASR model built on Google's EmbeddingGemma 2, is now on Hugging Face. It fine-tunes only the top 2 of 12 encoder layers (USM-style conformer) and trains a 6-layer Transformer reader with CTC — one forward pass, no autoregressive decoder.
- 3,700× faster than real time on a single H100; same weights decode in full-context or streaming mode (0.48s chunks)
- 6.87% average WER across 8 Vietnamese benchmarks, beating Gemini-Flash-3.8, Cohere Transcribe, and ElevenLabs Scribe v2
- Trained on 2,800 hours of public Vietnamese speech; lowercase, punctuation-free output with numbers as words
- Ships with pip install and one-line inference; requires accepting embeddinggemma-2 terms first
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