Pika launches Pika Music generation model
Video-generation startup Pika has released Pika Music, a single model that turns text, lyrics, voice samples and reference tracks into full songs, moving the company into the crowded AI music market.
Pika announced Pika Music, a music generation model that takes four kinds of input in one system: text prompts, lyrics, a voice reference and a music reference. Users can also combine them. Until now Pika has been known mainly for video generation.
The company says lyrics go into the model as conditioning tokens alongside the prompt, "so the words are part of the denoising objective rather than an alignment pass applied afterward." Section tags such as [verse] and [chorus] set the song's structure. Because the lyrics and the prompt are handled separately, the same words can be rendered in different styles.
A short vocal clip can guide the singer's timbre and delivery. Pika says the reference shapes the performance as it is generated, rather than a voice being laid over a separate backing track. Music references are separated and normalized first, so the new track picks up their texture and feel without being stitched together from parts of the source. For both voice and music references, users must confirm they hold the rights to what they upload.
On speed, Pika says Pika Music produces a 90-second song in 6.21 seconds on average in its local tests, about 14.5× faster than real time. The company also reports Audiobox Aesthetics scores close to Suno's, but on its own 15-prompt benchmark. Neither the speed nor the quality figures have been independently verified.
The announcement does not give pricing, availability details or the model's architecture.