Back to Blog

Say Hello to Pika Speech

Any voice from five seconds of reference audio. Up to five minutes of speech per request. Spoken before you finish reading this sentence.

August 18, 2026 • By Pika Team

Any voice from five seconds of reference audio. Up to five minutes of speech per request. Spoken before you finish reading this sentence.
Pika Speech is our text-to-speech model: a 3B flow-matching transformer that generates studio-quality 48 kHz speech in only a few denoising steps, clones a voice from a few seconds of reference audio, and speaks with controllable style, pace, and duration. For three-minute requests in our locally run tests, Pika Speech achieves a real-time factor of 0.02. It’s fast enough that typing the text takes longer than generating the speech. Try it on Pika API Club.
This kind of speed makes it possible for Pika Speech to be 9x more cost-efficient than ElevenLabs v3, 4.5x more cost-efficient than Cartesia and ElevenLabs Turbo, and 2x more cost-efficient than Fish Audio.* *Prices recorded as of 8.14.26

Hear it first

Voice clone

Voice reference
Pika Speech output

Preset voice

“Aahhh gas station sushi. a tuna roll purchased between the motor oil and the lottery tickets. Notes of regret. Hints of diesel. 4 Stars.”
 
"For this week only....... absolutely everything in store is up to 40 percent off."
 
“Dawn breaks over the open-plan savannah. Here, beneath the flickering fluorescent canopy, a colony of homo sapien-officius stirs — drawn, as they are each morning, toward the warm glow of the screen. But something is wrong. A scent drifts across the plain. David — a mid-ranking managerial male — has returned to the lunchroom and reheated his leftover fish pie. It is an act of extraordinary boldness. The herd freezes. Nostrils flare. Nobody will say anything. Nobody ever does.”
 

How it works

Flow-based speech generation

Pika Speech teacher is a latent flow-matching model. Text goes in through a large language model encoder; its hidden states are projected by learned connectors into conditioning tokens that carry both the transcript and the delivery caption. Audio lives in a compact latent space: a VAE compresses 48 kHz stereo into 25 latent frames per second, 128 channels each — a one-minute clip is just 1,500 tokens.
notion image
The generator is a 3B-parameter diffusion transformer (48 blocks) that learns the flow from noise to speech. During training we corrupt a clean latent toward noise along a straight path,
and train the model to predict the velocity that points back to the data. At inference we start from pure noise and integrate that flow; the result is a clean speech latent, decoded to a mel-spectrogram and rendered to a 48 kHz waveform by a neural vocoder with bandwidth extension.
Flow matching gives us the property that everything else in this post builds on: the mapping from noise to speech is smooth enough to be distilled into a handful of steps—eight or even fewer.

Speech DMD

Pika Speech is further distilled into a few-step generator by distribution matching distillation. Instead of teaching a student to imitate the teacher's trajectory step by step, DMD teaches it to match the teacher's output distribution: a frozen teacher score and a trained critic score together provide a gradient—evaluated directly in clean-speech space—that pulls the student's few-step samples toward the distribution the teacher would have produced. The student retains the teacher's voice, stylistic range, and duration behavior, but reaches the result in a few fixed steps, with guidance distilled into the model.
notion image

One anchor controls duration and pace

Ask a TTS system to say the same sentence in two seconds or in four, and you've asked for two genuinely different performances. Most systems control duration by trimming or time-stretching. Pika Speech controls it inside the model with a mechanism we call the EOS latent.
The idea: prepend a clean latent — never noised, never contributing to the loss — that carries what end of speech looks like, and give it the RoPE position of the frame where the utterance should land. The model literally sees where its sentence must end. Slide that anchor earlier and delivery compresses; slide it later and the same words relax and breathe. One knob, two effects: total duration and speaking rate.
notion image
Around the anchor sit two quieter mechanisms. The latent canvas itself is a duration prior — the model paces speech to fill the frames it's given, and a lightweight planner picks the frame count from the text at zero latency cost for plain requests. For the streaming variant of Pika Speech, a per-frame EOS head predicts "has speech ended?" so generation can stop the moment the sentence does, rather than at a preset buffer edge.

What the model learned from

The final training set contains 403,000 hours of filtered speech. It combines open corpora selected for breadth — including conversational, read, and expressive speech in English and Chinese — with a large-scale collection of in-the-wild audio. As we continue to refine and expand the models capabilities, we’ll add more language data.
Every recording passes through the same processing pipeline:
  • Voice-activity segmentation splits continuous recordings into clean utterances.
  • Denoising and loudness normalization produce a more consistent acoustic signal.
  • Quality filtering uses DNSMOS, aesthetic scoring, and dedicated audio-quality models to remove low-quality samples.
  • ASR transcription generates the text paired with each accepted utterance.
The remaining samples are converted into the same structured style-caption format used to control the model at inference. We also annotate end-of-speech boundaries at latent-frame resolution and pre-encode the audio into the latent representation used during training.

Inference acceleration

Few-step inference of a 3B model is already quick. It is not, by itself, why a minute of speech takes about a second. The serving stack is:
  1. FlashAttention-3 throughout the transformer. We use optimized attention kernels for both self-attention and cross-attention. Conditioning sequences are trimmed to their valid lengths, avoiding masked computation on unused tokens.
  1. Token packing. A long request is split at sentence boundaries, and all chunks are denoised together as one packed sequence with segment-isolated attention. Denoising latency grows sublinearly with the number of chunks: five chunks cost roughly twice as much as one, not five times as much.
  1. A compiled, full-precision vocoder. Optimizing the rest of the stack allows us to keep the waveform stage at full quality. For long requests, the vocoder is now the largest single cost, accounting for more than half of total latency.
  1. CUDA graphs for the full denoising loop. For each duration bucket, we capture all eight denoising steps as a single replayable graph at startup. Instead of launching thousands of kernels, a request replays one captured graph in ~0.1 s.
  1. Fused RoPE and packed graphs. Pika Speech computes rotary positional embeddings in a single fused kernel and extends graph capture to the packed multi-chunk path. Together, these optimizations reduce one-minute generation time from 1.29 s to 1.04 s.
notion image

Benchmarks

Speed comparison with efficient models

Speed is measured on a long-form corpus built for this evaluation: 200 narrations—100 English and 100 Chinese—ranging from one to five minutes and covering twenty distinct reference speakers per language. We report real-time factor (RTF), defined as generation time divided by the duration of the generated audio. An RTF of 0.02 means that one minute of speech takes about 1.2 seconds to generate. Each point in the figure reports the median for its duration bucket.
notion image

Quality comparison with leading TTS systems

We evaluate English and Chinese voice-cloning quality on 2,000 samples per language. For each sample, a system receives a reference recording and a new transcript, then synthesizes the transcript in the reference speaker's voice. The metrics cover three aspects of output quality: intelligibility, speaker similarity, and perceptual audio quality.
  • WER ↓ (English) and CER ↓ (Chinese) measure transcription errors in the generated speech relative to the target transcript. Lower scores indicate fewer content errors.
  • Resemblyzer similarity ↑ measures the similarity between speaker embeddings from the generated and reference audio. Higher scores indicate a closer match to the reference voice.
  • DNSMOS OVR ↑ estimates overall perceptual speech quality.
  • AudioBox PQ ↑ provides a model-based estimate of perceptual audio quality.
Model
Price / min
En WER % ↓
En Resem↑
Zh CER % ↓
Zh Resem↑
En DNSMOS ↑
Zh DNSMOS ↑
En PQ ↑
Zh PQ ↑
ElevenLabs v3
$0.09
1.879
80.88
1.450
78.90
2.912
3.042
6.551
7.143
MiniMax Speech 2.8 HD
$0.051
1.691
90.16
0.846
90.93
3.129
3.297
6.296
6.788
Cartesia Sonic 3.5
$0.045
1.583
87.41%
1.208%
88.03%
3.1595
3.3273
6.872
7.130
seed-audio-1.0
$0.15
1.625
89.27
3.201
89.13
3.183
3.269
6.137
6.745
Fish Audio S2.1 Pro
$0.21
1.727
83.74
1.155
80.49
3.165
3.210
6.236
6.959
Pika Speech
$0.01
1.990
80.30
1.727
75.41
3.188
3.247
6.772
6.957
Table note: ↓ indicates lower is better; ↑ indicates higher is better. Among the systems shown, green highlights the best result in each column and yellow highlights the second-best. Tied results receive the same highlight.

Future work

Pika Speech is a foundation for a broader real-time generation stack. We plan to continue improving speech quality, controllability, long-form consistency, and streaming latency. These advances will also support PikaStream 2.0, our real-time video generation model, by providing low-latency speech for synchronized audiovisual generation.

Availability

Pika Speech is live now in the Pika API Club. Try it out and let us know what you think.