Pika launches four low-cost audio models
Pika has released four audio models covering video-to-audio, music, sound effects and speech, and says they cost up to 20 times less than rivals, though it has published only its own benchmarks.
Generative-video startup Pika announced Pika Audio Models, which it calls its first family of foundation sound models. The four models are available now, only through the Pika API Club.
The lineup:
- Pika Soundtrack turns video into synchronized music, speech, ambience and motion-aware sound effects. It can run with a blank prompt or take direction on what to include or leave out.
- Pika Music takes text prompts, lyrics, vocal references and reference tracks, alone or combined, and produces songs up to six minutes long. Pika says it generated a 90-second song in 6.21 seconds on average.
- Pika SFX turns text into up to 20 seconds of 44.1 kHz stereo audio. Pika reports an average generation time of 0.847 seconds in a local benchmark.
- Pika Speech is a text-to-speech model with preset voices and voice cloning from a few seconds of reference audio. It outputs 48 kHz audio and accepts requests up to five minutes long.
Pika says efficient training, few-step generation and a speed-focused inference stack make these the lowest-priced audio models on the market. It claims Speech is 9x more cost-efficient than ElevenLabs v3 and 2x more than Fish Audio. It says Soundtrack is 2x more cost-efficient than Hunyuan Foley.
In Pika's own full-duration test against LTX-2.3 Foley V2A, HunyuanVideo-Foley and MMAudio v2, Soundtrack had the strongest semantic alignment and the lowest audio-visual desynchronization. It ran at 0.617 seconds of wall time per generated second of audio.
All benchmarks were run by Pika. The post gives no absolute prices, no evaluation details and no information on model sizes or training data.