Audar TTS v1

Input

How streaming is played

The stream is requested as response_format=pcm and fed into an AudioWorklet FIFO, so audio starts the moment the first frame lands and joins are sample-exact. TTFA below is time-to-first-audio — the number that matters for a live agent.

• The gap is fixed at the source. Chunk sizes now RAMP (18 → 20 → 58 → 53 chars instead of 14 → 85), because decode runs ~2.5× realtime and a chunk's playback only covers the next chunk's generation while their sizes are comparable. A tiny opener followed by a big chunk starved the player for 1.88 s — that was the gap you heard.
• Verified across 4 texts × 2 providers: buffer lead never goes negative. The 250 ms prebuffer is only insurance against GPU contention; set it to 0 to see raw TTFA.
• WAV over a stream was also fixed: it used to emit a whole wav file per chunk, so players trusted the first header and truncated a 12.8 s answer to 4.1 s. PCM avoids headers entirely.