AIPhoton adds streaming speech with Qwen3-TTS and Kokoro
The announcement claims first audible audio in under 85 milliseconds at the 95th percentile, 56–70% sooner than vLLM-Omni on H100 and B200.
TLDR
Photon now supports Qwen3-TTS in 0.6B and 1.7B versions, plus Kokoro-82M, with streaming output, according to the announcement.
The author reports first audible audio in under 85 milliseconds at the 95th percentile—56–70% sooner than vLLM-Omni on H100 and B200. They say the measurements used six requests per second, identical arrival schedules for both engines, and vLLM-Omni's tuned chunk-ramp configuration. The metric measures the first audible audio, not the first chunk, which can be silence.
In follow-up replies, the author says setting stream=True delivers audio piece by piece, allowing playback before the sentence is finished. Kokoro has 82 million parameters and runs on CPU, but its Photon implementation takes phonemes as input rather than text.
Photon adds streaming speech with Qwen3-TTS and Kokoro
The announcement claims first audible audio in under 85 milliseconds at the 95th percentile, 56–70% sooner than vLLM-Omni on H100 and B200.
TLDR
Photon now supports Qwen3-TTS in 0.6B and 1.7B versions, plus Kokoro-82M, with streaming output, according to the announcement.
The author reports first audible audio in under 85 milliseconds at the 95th percentile—56–70% sooner than vLLM-Omni on H100 and B200. They say the measurements used six requests per second, identical arrival schedules for both engines, and vLLM-Omni's tuned chunk-ramp configuration. The metric measures the first audible audio, not the first chunk, which can be silence.
In follow-up replies, the author says setting stream=True delivers audio piece by piece, allowing playback before the sentence is finished. Kokoro has 82 million parameters and runs on CPU, but its Photon implementation takes phonemes as input rather than text.
