NVIDIA, Mistral, and Phonic Release Voice AI Models, Top Benchmarks

Phonic, a ten-person startup, released Phonic-1, a native speech-to-speech model that ranks first on Artificial Analysis's speech-to-speech index with a score of 87.1, narrowly ahead of xAI's Grok Voice 3.0 at 86.9. The model can initiate tool calls mid-sentence; in a demo it completed five lookups while the caller was still speaking and responded 1.1 seconds after the caller stopped, versus 15.7 seconds for OpenAI's GPT-Live. Phonic-1 completed 79.9% of customer-service tasks with a 0.61-second time-to-first-audio latency. On Cekura's benchmark across 12 real-time voice models, Phonic-1 passed 100% of calls in categories including plain booking and lookup, tool and system errors, background noise, and safety calls, and saved caller data correctly on 97.5% of relevant calls, ranking first. Its weakest area was interruption handling at 87% passing, against a 90% median. Phonic raised a $27.5M Series A led by First Round Capital, bringing total funding to $31.5M, and the model powers Ashby's AI job interviews across 25 countries. Phonic charges by the minute with no LLM pass-throughs or per-token charges.

NVIDIA released Nemotron 3.5 Transcribe, a single open-weight model handling transcription, speaker diarization, and translation across 40 languages through plain-language prompts. It ranks first on the Hugging Face Open ASR benchmark and second among open-weight models on Artificial Analysis. With approximately 4.3 billion active parameters, NVIDIA says it uses 80% fewer parameters than the top open-weight competitor. It is available in BF16, FP8, and NVFP4 formats and supports RTX Spark, DGX Spark, Jetson AGX Thor, and GeForce RTX 5090. A companion release, Granary v2, targets adaptable multilingual transcription use cases.

Mistral's Voxtral Mini Transcribe 3 placed second out of nine models on Voice Arena's Diarization and ASR leaderboard with a 40.4% DER, outperforming its larger sibling Voxtral Small Transcribe 3, which scored 41.1% DER and placed fourth. Both were evaluated on 139 held-out conversations with human-verified references, a strict 0 ms collar, and no predefined speaker counts. Missed speech accounts for 82% of Mini's total DER; both models correctly identify the exact speaker count in 83.5% of sessions.

Deepgram agreed to acquire Daily, the team behind open-source voice-agent framework Pipecat, and committed to keeping Pipecat free and open source following the acquisition.