Pika launches sub-second sound effects model
Pika has released Pika SFX, a text-to-sound-effects model that its own benchmark puts at under a second per clip, aimed at interactive tools that need sound effects generated on demand.
AI video startup Pika announced Pika SFX, a model that turns a written description into a sound effect in real time. Prompts can describe a single event, such as glass shattering, or a sequence with timing, environment, texture and mood. Pika says the model leaves out speech, music and background noise unless the prompt asks for them.
The model outputs 44.1 kHz stereo audio. A transformer text encoder turns the prompt into conditioning embeddings. A text-conditioned diffusion transformer (DiT) then generates the sound in a compressed acoustic latent space, and a "semantic-acoustic autoencoder" decodes it into a waveform.
Pika trained the model in three stages. The first used flow matching. The second used teacher–student distillation to cut the number of inference steps. The third was post-training on curated data with human feedback. The company also says it tuned the model, hardware and serving stack together.
Pika's benchmark used 50 prompts, with 454 seconds of requested audio per model and clip lengths from 2 to 20 seconds. On average, Pika SFX finished end to end in 0.847 seconds, including CPU processing and saving the file. The fastest competitors were ElevenLabs v2 at 2.47 seconds and Stable Audio 3 SFX at 2.64 seconds. Pika ran its own model on local hardware but measured the others through their hosted APIs, so the numbers are not directly comparable. Pika ran the benchmark itself.
The source cuts off at the start of the audio-quality results. It gives no quality scores, pricing or availability details.