Pika launches Soundtrack video-to-audio model
Pika has released Soundtrack, a model that generates synchronized sound effects, voice, music and ambience for an entire video, which puts the generative video startup into the video-to-audio market.
Pika announced Pika Soundtrack, a model that takes a video and generates a new audio track covering the full scene. Users can add an optional natural-language instruction to tell the model what to emphasize, include or leave out, for example "no dialogue / no music."
Pika says the core model is a transformer-based latent diffusion model. Video is compressed into visual latent tokens that keep timing, motion and scene layout. The text prompt is encoded into semantic tokens. Audio is generated in a compressed latent space rather than as raw waveform samples. At each denoising step, the audio attends to both the video and text tokens, so the model can decide what sound to make and when. Training uses a flow-matching, diffusion-style denoising loss, with the video latent fixed as conditioning.
To handle long videos, Pika says it uses distillation to cut the number of generation steps, caches visual and temporal features shared by neighboring chunks, and uses sparse attention focused on the local audio and video regions that drive synchronization.
Pika also reports benchmark results. It ran its evaluation in temporal chunks across the full length of each video instead of scoring only the first few seconds. By its account, Soundtrack has the strongest semantic alignment, the lowest audiovisual desynchronization and the fastest full-duration generation of the systems it tested. The company also says the model is up to 2× more cost-efficient than competing video-to-audio models, based on local evaluation.
These benchmarks are self-reported, and the excerpt available does not name the competing models or give the full scores. It also does not cover pricing or availability.