Phonic, NVIDIA, and Mistral voice AI

NVIDIA released Nemotron 3.5 Transcribe on October 28, a single open-weight model that handles transcription, speaker diarization, and translation across 40 languages through a promptable interface. NVIDIA states it ranks first on the Hugging Face Open ASR benchmark and second among open-weight models on Artificial Analysis, using approximately 4.3 billion active parameters, described as 80% fewer than the top open-weight competitor. The model is available in BF16, FP8, and NVFP4 formats and supports RTX Spark, DGX Spark, Jetson AGX Thor, and GeForce RTX 5090 hardware. It can also answer questions about audio content.

Phonic released Phonic-1 on October 29 as a native speech-to-speech model, alongside a $27.5M Series A led by First Round Capital, bringing total funding to $31.5M. On Artificial Analysis's speech-to-speech index it scored 87.1, above xAI's Grok Voice 3.0 at 86.9. Phonic states it completed 79.9% of customer-service tasks with a 0.61-second time-to-first-audio latency. On Cekura's 12-model benchmark, Phonic-1 saved caller data correctly on 97.5% of relevant calls, ranking first, and passed 100% of calls in categories including plain booking and lookup, tool and system errors, background noise, and safety and scope calls. On Medicare compliance and routing calls, Phonic-1 achieved a 92% pass rate versus a median of 49% across models. Interruption handling was Phonic-1's weakest area, at 87% passing against a 90% median. Phonic charges by the minute with no LLM pass-throughs or per-token charges. A discrepancy exists between Phonic's self-reported Cekura score of 90.2 and Cekura's public listing of 68.3% for Phonic v1; the available items do not explain it.