In late October 2026, several key developments in speech and audio AI models were reported across the industry.
Voice Cloning and TTS Releases
Edge0 released Audio8-TTS-Realistic, a 0.6B voice cloning Text-to-Speech model (Evidence 1). Simultaneously, AssemblyAI shipped AssemblyAI-TTS-1, positioned as a $15 per 1M character Text-to-Speech API tailored for voice agents (Evidence 6).
Full-Duplex and Multilingual Speech
Block unveiled Dessa, a full-duplex speech model trained exclusively on speech data (Evidence 5). Expanding multilingual capabilities, NVIDIA released Nemotron 3.5 Transcribe, an open 30B Mixture-of-Experts (MoE) model designed for 40-language transcription featuring approximately 4B active parameters (Evidence 7).
*Analysis:* Based strictly on the published announcements, these releases highlight a continued diversification in audio AI architectures—ranging from compact voice-cloning models and specialized APIs to full-duplex systems and large-scale open MoE transcription frameworks. No hands-on testing or independent benchmarking was conducted to verify these reported capabilities.