🎙️ VoxCPM Streaming Demo

Windowed streaming TTS with chunk-to-chunk voice cloning, powered by openbmb/VoxCPM2 via the public API of openbmb/VoxCPM-Demo. Text is split on sentence boundaries; chunk 0 establishes the voice and every later chunk clones it, so audio starts playing after the first short sentence instead of the whole passage.

Input

Text to synthesize

Voice description — applied to chunk 0 only; the clone carries it onward

🎤 Voice cloning & advanced settings

Reference audio — optional; seeds chunk 0's voice. 5–20 s, max 50 s

Reference transcript

Chunk size (characters) 120

Buffer ahead before playback

CFG guidance 2.0

Normalize text (numbers, dates, abbreviations)

Denoise reference audio

Single-shot mode (no splitting — one call for the whole text)

Playback & metrics

TTFB

—

RTF

—

Underruns

—

Measured upstream: a short sentence takes ~3–7 s — slower than real time (RTF ≈ 1.5), so expect occasional gaps between chunks. Windowing still cuts time-to-first-audio to one sentence instead of the whole passage, which is the point. Truly gapless streaming needs RTF < 1.0, i.e. a dedicated GPU backend. Raise Buffer ahead to trade latency for smoothness.

Metrics
TTFB — wall time until the first audio starts playing.
RTF — generate time ÷ audio duration. Below 1.0 means generation outruns playback.
Underruns — times the player ran dry waiting for the next chunk, i.e. audible gaps.

Static Space · no model hosted here · calls the public Gradio API of openbmb/VoxCPM-Demo. Windowing mirrors the /tts_stream design: sentence-bounded chunks, chunk 0 as the clone prompt.