Meta and Cartesia each claim top streaming voice AI rankings
Streaming speech-to-text at low latency is a building block for voice agent systems, and multiple providers are posting competing benchmark claims in this space. Latency at each step in the voice stack is perceptible to users, making benchmark positioning significant for developers choosing providers.
On September 7, 2026, Cartesia released Sonic-3.5/3.6 for text-to-speech with 90ms latency and Ink-2 for speech-to-text with 100ms latency, with Cartesia claiming to hold the top streaming ranking for both directions of voice AI simultaneously. Meta Superintelligence Labs released Muse Voice Transcribe on September 1, processing audio in 80ms chunks with a reported 3.1% final-transcription word error rate and ranking first on Artificial Analysis streaming speech-to-text and public diarization benchmarks. The model handles speaker diarization for more than 20 speakers, turn-endpointing, and multilingual code-switching natively in a single model. Meta distributes it through the Meta Model API, Meta AI for Mac, and Muse Code. Meta released Muse Spark 1.3 on September 4 with a max reasoning tier delivering stronger coding and agentic performance than the previously available high and xhigh tiers.
Cartesia released Sonic-3.5/3.6 for text-to-speech at 90ms latency and Ink-2 for speech-to-text at 100ms latency on September 7, claiming to be 'the only provider to have #1 models for both speaking and listening'.
That puts Cartesia in direct competition with Muse Voice Transcribe, Meta's real-time transcription model released September 1, which holds first-place rankings on Artificial Analysis streaming speech-to-text and public diarization benchmarks.
Meta released Muse Spark 1.3 with a new max reasoning tier on Muse Code and the Meta Model API
Meta Superintelligence Labs released Muse Voice Transcribe on September 1, its first real-time audio perception model, handling streaming speech-to-text, speaker diarization for more than 20 speakers, and turn-endpointing within a single model.
The model processes audio in 80ms chunks and Meta reports a 3.1% final-transcription word error rate and first-place rankings on Artificial Analysis streaming speech-to-text and public diarization benchmarks. It is available through the Meta Model API, Meta AI for Mac, and Muse Code.
I send one email each morning with the stories that moved. If you would rather just read here, that works too.
Subscribe free