Streaming latency

Client TTFA starts before the HTTP request and ends at the first decoded nonempty SSE audio chunk. Standard Qwen emits one buffered recording; this is not incremental synthesis. SSE transport excludes browser playback buffering.

Run Engine Checkpoint Incremental Calls Realtime speed (×) Median TTFA ms p95 TTFA ms Median RTF p95 RTF Errors
20261025-050404-08222176 faster 0.6B-CustomVoice True 100 2.79 303.8 353.6 0.356 0.396 0
20261025-050404-08222176 faster 0.6B-Base True 30 2.89 312.5 364.4 0.347 0.361 0
20261025-053632-9d5b8428 faster 1.7B-CustomVoice True 100 2.51 327.2 375.1 0.397 0.430 0
20261025-053632-9d5b8428 faster 1.7B-Base True 30 2.61 341.7 386.8 0.382 0.407 0
20261025-053632-9d5b8428 faster 1.7B-VoiceDesign True 100 2.46 328.3 372.8 0.406 0.438 0