Voxtral 3

Commenters discussed Mistral’s Voxtral 3 release as a combined STT/TTS/agent system and debated its realism, openness, and practicality. Some praised transcription and diarization quality for podcast and personal use and reported success running smaller variants (Voxtral Mini/other local models) on commodity hardware, while others said the new TTS sounds more artificial than earlier versions. Several people flagged the demo as unrealistic (the dermatologist appointment example) and noted quirky behavior like accent shifts. Many recommended alternative local ASR/TTS options (Kokoro, Qwen ASR, Gradium, pocket-tts) and tools for subtitle syncing, and asked about latency, full‑duplex interaction, and integration details.

Opinion divided sharply over deployment and safety. Numerous commenters rejected the closed model and worried about speech data leaving local control, while others welcomed hosted EU API or self‑deployment options. Some argued Voxtral‑style agents could improve customer experience by answering calls quickly; others warned the same capabilities make fraud and social engineering easier if agents are granted intranet or payment powers. Technical debates focused on real‑time diarization, context biasing, comparisons to Gemini Live and Nvidia Nemotron, and whether self‑hosting smaller models will satisfy privacy and performance needs.