A Speaches sentence is said clause by clause, so its first audio comes after its first clause

Shipped

October 28, 2026 at 5:38 AM UTC

Author

Kamo

Commit

7611640

Luna), but Speaches' Kokoro/Piper make each request whole before the first byte (~0.6 s + 25 ms a character on k3m1), so an SI with an in-cluster voice (sp00verify's Nell, TPEARDEAF: 0.9-1.9 s) waited for the whole sentence. A streamed PCM request to Speaches is now cut after its commas, semicolons, colons and dashes (never inside 1,000; tiny clauses join a neighbour; at most four pieces), the first clause is handed on as soon as it is made and the next is made beside it (at most two in flight), and the silence at each join is trimmed to about a comma's pause. An empty or failed first clause still fails over; other formats and providers are asked once, as before.