Pika launches 3B-parameter Pika Speech TTS model
Video-generation startup Pika has released Pika Speech, a text-to-speech model built for speed and low cost, with a real-time factor of 0.02 in the company's own local tests.
Pika announced Pika Speech, a 3B-parameter flow-matching transformer that produces 48 kHz speech in a few denoising steps. It can clone a voice from a few seconds of reference audio and lets users control style, pace and duration. It is available through Pika API Club.
Pika says that in its local tests on three-minute requests, the model reached a real-time factor of 0.02, or about one second of compute per minute of audio. Using prices recorded on August 14, 2026, the company says this makes Pika Speech 9x more cost-efficient than ElevenLabs v3, 4.5x more than Cartesia and ElevenLabs Turbo, and 2x more than Fish Audio.
The architecture runs text through a large language model encoder. A VAE compresses 48 kHz stereo audio into 25 latent frames per second, so a one-minute clip comes to 1,500 tokens. A 48-block diffusion transformer then generates the audio. Pika used distribution matching distillation (DMD) to turn the teacher model into a student that needs eight or fewer steps.
Pika says the main technical change is an "EOS latent," a clean anchor frame placed at the position where the utterance should end. Moving the anchor earlier or later changes both total duration and speaking rate, with no trimming or time-stretching. A streaming version adds a per-frame end-of-speech detector.
The model was trained on 403,000 hours of filtered speech, mostly English and Chinese. On the serving side, Pika cites FlashAttention-3 and sentence-level token packing. All of the performance figures come from Pika, and the company has not published third-party quality comparisons.