Gradium boosts speech-to-text accuracy on rare names with keyword prompting
mattturck · x · 2026-07-23
- Gradium says speech-to-text systems often fail on names because proper nouns are rare tokens that general models have seen very little.
- Its approach is keyword boosting: provide the names that matter for a given request so the STT model biases toward them.
- The company says it tested the feature on people, places, and organizations, and the example screenshot shows live transcription correctly handling names such as Vinčić, Mbappé, and MrBeast.
More from Models
- Gemini 3.6 Flash Becomes the Default Model for Managed Agents — _philschmid · 2026-07-23
- Cohere Confirms Community Quants for Arabic Speech Model on Hugging Face — cohere · 2026-07-23
- Reddit jokes about “obliterating” Kimi K3 the moment it ships — JsonBasedman · 2026-07-23
- Leak: OpenAI Targets September Release for GPT-6 as Autonomous AI Researcher — VraserX · 2026-07-23
- Rumor says GPT-6 has been delayed by several more months — patience_cave · 2026-07-23
- Moonshot’s Kimi K3 is blamed for Friday panic as AI prices keep falling — TiernanRayTech · 2026-07-23