Back to Blog
Introducing Pika Music
One model, many ways to make a song.
August 18, 2026 • By Pika Team
One model, many ways to make a song.
Start with an idea, finished lyrics, a reference voice, an existing track—or combine them—and generate a complete piece of music through one workflow.
Today we’re introducing Pika Music, a new music generation model for turning creative direction into finished tracks. It supports multiple inputs: text, lyrics, voice-condition, music reference, and combinations of those inputs in a single system.
Most music tools make creators choose a separate workflow for every starting point. We wanted the model to meet people wherever an idea begins. Sometimes that’s a sentence: dreamy synth-pop with intimate vocals and a euphoric final chorus. Sometimes it is a complete lyric sheet. Sometimes the important ingredient is a voice, a melody, or a song that needs to be reimagined.
Pika Music treats each of these as a different way to describe the same destination: the song you want to hear.
It is also built for iteration. In our local tests, Pika Music generates a 90-second song in 6.21 seconds locally on average—about 14.5× faster than playback—so creators can explore a new lyric, voice, or arrangement without leaving the creative loop.
Four input modalities
1. Prompt—go from description to a complete track
Text-to-music begins with ordinary creative direction. A prompt can describe genre, mood, instrumentation, vocal character, arrangement, or a progression across the song. The model turns those instructions into a coherent track rather than a short musical texture.
Prompt
modern pop R&B, B flat minor, 69 bpm, softly swung groove, mellow lofi textures, deep piano melody, vinyl crackle, sparse synth pads, warm low bass, clean reverbed female lead vocal, soulful lead with breathy ad-libs, subtle low male doubles on the chorus only, chill spacious mix
lyrics:
[verse]
Night after night I keep the light on low
Counting all the things I said I'd let go
[chorus]
So I stay, so I stay awake
Some kind of quiet only midnight makes
2. Lyrics—start with the words
Lyrics enter 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] are read as structure, informing phrasing and section boundaries, while the prompt conditions the arrangement around them.
Because the two paths are separate, the lyric can be held fixed while the prompt varies: the same words render as an acoustic ballad, a dance-pop single, or a cinematic rock performance.
Lyrics + direction
Text-to-Music Example 01 R&B Lo-Fi
Prompt
Modern pop R&B, B flat minor, 69 bpm, softly swung groove, mellow lofi textures, deep piano melody, vinyl crackle, sparse synth pads, warm low bass, clean reverbed female lead vocal, soulful lead with breathy ad-libs, subtle low male doubles on the chorus only, chill spacious mix
Lyrics
[Verse] Night after night I keep the light on low Counting all the things I said I'd let go [Chorus] So I stay, so I stay awake Some kind of quiet only midnight makes
Text-to-Music Example 02 Echoes of Goodbye
Prompt
Haunting emotional ballad steeped in longing, quiet resilience and heartbreak, soft and sorrowful mood like walking through empty rooms filled with memories, gentle melody carried by piano chords and orchestral swells that rise and fall like tides, fragile tender reverent delivery, spacious and cinematic.
Duration: 60
Lyrics
[Verse 1] I still hear your voice in the hush of the dawn, A whisper that fades, but never is gone. The bed still holds your shape like a ghost in the light, And I reach out to shadows in the heart of the night. [Pre-Chorus] We said forever in the breath of a kiss, But forever can crumble like this… [Chorus] You're the echo in my silence, the tear I cannot cry, A dream I keep reliving though I know it's goodbye.
3. Voice condition—give your song a singer
A short vocal reference can guide vocal texture, timbre, delivery, and energy. It can be used on its own or combined with a prompt and lyrics, allowing the musical arrangement and the vocal character to be directed together.
This is designed for creative voice conditioning—not merely attaching a voice to an independently generated backing track. The reference participates in the generation of the performance.
Note that our model requires user acknowledgement that they have the rights to use whichever voice they input.
Input · voice reference
Output
4. Music Reference—add the vibe
Some things about a song are easier to point at than to describe. A music reference lets you supply a track whose feel you are after, instead of trying to name the genre, instrumentation, and production character in words.
The reference shapes the musical world of the new song—its texture, instrumentation, energy, and overall feel—while the prompt and lyrics decide what the song actually says. The reference sets the vibe; the written direction sets the content.
Before generation, the reference is separated and normalized so its musical qualities can be read independently of the original mix. The new track is then generated as a single performance conditioned on those qualities, rather than stitched together from pieces of the source.
Note that our model requires user acknowledgement that they have the rights to use whichever track they input.
One model, multiple kinds of control
Under the hood, these workflows are different combinations of conditioning signals:
- Language describes genre, mood, instrumentation, vocal style, and arrangement.
- Lyrics specify the words and can provide high-level song structure.
- Voice references guide vocal character and delivery.
- Music references provide source material for a new vibe.
- Duration and seed provide additional generation control.
The challenge is not simply accepting more inputs. Those inputs can disagree. A lyric implies one structure while a reference song implies another; a requested style can pull against the source arrangement; vocal identity and intelligibility can compete. The system must resolve these signals into one musically coherent performance.
Our generation pipeline prepares each signal for the model. Text and lyrics are normalized into clearer creative instructions. Voice references are analyzed for useful vocal attributes, and source audio is separated and normalized before cover generation. The model then produces the complete track from the resulting conditioning.
The practical benefit is composability. Creators do not need a different model or interface every time the starting material changes.
Architecture
At its core, Pika Music is a latent diffusion transformer with a mixture of local and full attention layers. The model works in a compressed acoustic representation where long-range musical structure is computationally manageable, then decodes the finished latent sequence into a stereo waveform.
Different inputs enter through specialized conditioning paths in the form of conditioning tokens or acoustic context latents.
These signals meet inside one shared generative decoder. That design matters: a cover with a new voice and a new style is not assembled by chaining several independently generated outputs. The model resolves the source melody, lyrics, timbre, and written direction together while generating the target song.
This shared latent interface is what lets one model move from unconstrained composition to tightly conditioned transformation without changing the core architecture.
Training objective
We train the model using a flow-matching objective. For a clean target latent , Gaussian noise , and sampled noise level , we construct:
The transformer predicts the velocity that points from the clean example toward the sampled noise:
Here, represents whichever conditioning is available for the example: text, lyrics, voice, or source-song context. During training, conditioning is selectively perturbed or dropped. This teaches the model to remain useful when references are imperfect and enables classifier-free guidance at inference time.
The joint objective serves two purposes. It teaches the model how to use each new conditioning signal, and it preserves the general composition ability of the model.
Benchmark evaluation
We built a fixed-lyrics benchmark to compare music-generation systems under the same creative directions, rather than selecting only favorable samples. The full benchmark contains 15 prompts spanning acoustic folk, pop, R&B, rap, rock, country, electronic music, metal, jazz, cinematic music, children’s music, Afrobeats, and multilingual songs. Every prompt specifies the style, lyrics, requested duration, language, BPM, key, and time signature.
Each system generates two outputs per prompt, producing 30 outputs per measured system. We evaluate two complementary dimensions:
- Music quality: Audiobox Aesthetics measures enjoyment, usefulness, production complexity, and production quality. SongEval measures coherence, musicality, memorability, clarity, and naturalness. MuQ-MuLan measures prompt alignment and within-prompt diversity. We also estimate tempo and key to evaluate musical control.
- Latency: We measure how long it takes for a generation to become usable. Hosted providers are timed end to end, from submitting a request through downloading the resulting audio.
The benchmark evaluates the text/lyrics-to-music path.
Results
The headline results are summarized below. Automatic evaluators are useful scalable proxies, but small differences should not be interpreted as definitive human-preference rankings.
Generation performance
Music generation is iterative. People adjust a lyric, change the singer’s energy, try another arrangement, or explore several directions before choosing one. Latency therefore changes the creative experience: every long wait interrupts the loop between an idea and hearing the result.
We optimized the complete inference path for a concrete model, hardware, and serving configuration. The result is a warm generation path that is substantially faster than real time.
In our synchronized production benchmark, the optimized system generated a 90-second song in 6.21 seconds on average. Across all 30 measured generations, mean latency was 6.25 seconds, median latency was 6.14 seconds, and P90 latency was 8.11 seconds. The median real-time factor was 0.068, or approximately 14.6× faster than playback.
Figure: End-to-end music-generation latency on a logarithmic scale. Diamonds show means, hollow points show medians, vertical caps show P90, and bars show the observed min–max across 30 generations.
Music quality
No single score captures whether a song works. We evaluate complementary properties—technical quality, musical coherence, prompt alignment, memorability, vocal naturalness, diversity, tempo, and key—and pair automated evaluation with listening.
On the text/lyrics to music benchmark, Pika Music reached the top tier of the Audiobox Aesthetics evaluation. Its Content Enjoyment score was 7.403, compared with 7.412 for Suno, and its Production Quality score was 8.064, compared with 8.151 for Suno.
Pika Music combines competitive automatic aesthetics with a broad, composable workflow and a substantially faster measured inference path.
Make music from wherever the idea begins
A prompt, finished lyrics, a reference voice, or an existing song can each be the beginning of a track. Pika Music brings those inputs into one model, so creators can explore, combine, and revise ideas without switching between separate tools.
Start in the playground to try Pika Music directly. When you are ready to bring it into your own product or workflow, use the API page to start building.