Speech vendors never lack “faster and cleaner” posters. Meta’s September 1 drop, Muse Voice Transcribe, is sharper: it welds three jobs that usually live in separate pipes—streaming ASR, diarization, and endpointing—back into one autoregressive token stream.
The technical story is concrete. The model sits in the Muse Spark family. Audio arrives in 80 ms chunks (12.5 Hz) as soft tokens; each step either keeps listening (<|next_audio|>) or emits text. When the stream ends, <|empty_audio|> flushes remaining tokens. Latency is not a fixed knob—reinforcement learning trades word-error rate against delay per word (“adaptive delay”). Speaker changes fire <|start_of_turn|>; identities land as <|speaker_A|>-style tags; onset and endpoint get their own symbols. Meta says hour-plus audio and 20+ speakers need no separate post-processing pipeline.
Benchmarks and price decide procurement. Meta claims first place on Artificial Analysis’s streaming speech-to-text index and on public diarization boards as of 2026-09-01. VentureBeat’s AA table puts final streaming WER near 3.1%, ahead of Cartesia Ink-2 (3.4%), ElevenLabs Scribe v2 Realtime (3.6%), and GPT Live Transcribe / Grok Speech (~3.9%). The list price is the real knife: $0.18 per audio hour, streaming and batch at parity—an order of magnitude under OpenAI’s streaming row in the same comparison. Training covers 70+ languages with 25 validated at launch, plus in-sentence and cross-sentence code-switching and keyword/context biasing. Distribution: Meta Model API, Meta AI for Mac, and Fn-hold dictation inside Muse Code.
Take: this is not “another ASR.” It promotes the two things that actually break meetings, contact centers, and voice agents—who is speaking, and whether the turn is over—from plugins into jointly trained primary tasks. Eighteen cents an hour will force rivals to rewrite pricing pages; concurrent-stream and starts-per-hour caps (developer drafts have floated figures like 8 concurrent streams and 1,000 starts per hour) will decide whether call-center load survives contact. Enterprises should watch three numbers: real WER in noisy rooms and dialects, speaker-tag stability under crosstalk, and whether ZDR truly prices at parity. The voice-agent stack is shifting from “transcribe, then orchestrate” to “perception that already understands turns.” Meta just swung both the price knife and the architecture knife.
[1][2]