Rápido TTS

Tiny, fast Brazilian Portuguese text-to-speech. 21M parameters, a 42 MB checkpoint, 48 kHz audio and Apache 2.0. It runs offline on your own GPU or CPU, so no text or audio leaves the machine.

• Reads Brazil right: numbers, money, dates, times, ordinals, units and phone numbers come out the way a Brazilian says them, with 2.94% WER on our numbers & pronunciation test. That is the lowest of the small open pt-BR TTS models we tested.
• Small: 21M parameters and a 42 MB checkpoint, trained from scratch for Brazilian Portuguese.
• Fast: 76× real time on one GPU, with first audio in 129 ms and only 1.2 GB of GPU memory.
• Private: it runs fully offline on your GPU or CPU.

Roadmap

Rápido v1 is the first of a family of small, fast speech models.

• Rápido v1: Brazilian Portuguese TTS, 21M parameters, one built-in voice
• Rápido Clone: zero-shot voice cloning from a few seconds of reference audio
• Rápido Multilingual: one compact model for many languages

Code: github.com/anurag12-webster/rapido-tts

Listen

"Olá! Tudo bem com você? Hoje o dia está lindo em São Paulo."

"O boleto de R$ 1.299,90 vence no dia 05/11/2026, às 14h30."

"Você vem com a gente ou vai ficar em casa? Já são quase oito horas!"

Each clip was made with the default settings (8 steps, guidance 4.0, seed 0), straight from the text shown.

Install

git clone https://github.com/anurag12-webster/rapido-tts && cd rapido-tts
pip install -r requirements.txt

Requires Python 3.10+. For a smaller CPU-only install, install PyTorch first: pip install torch --index-url https://download.pytorch.org/whl/cpu, then pip install -r requirements.txt.

Use

from rapido import Rapido

tts = Rapido.from_pretrained()                 # downloads the weights once; GPU if available

result = tts.synthesize("Olá! Tudo bem com você?")
result.save("ola.wav")

synthesize() returns a Result:

• result.output["audio"] is float32 mono PCM.
• result.output["sample_rate"] is 48000.
• result.output["text"] is the spoken form the model read.
• result.metrics holds the audio length, the time taken and the speed relative to real time.

Streaming

for chunk in tts.synthesize(long_text, stream=True):
    play(chunk.output["audio"], chunk.output["sample_rate"])

Each chunk is one sentence, returned as soon as it is ready. Joined together, the chunks are the full text.

Settings

from rapido import Rapido, Config

tts = Rapido(Config(device="cpu", steps=8, guidance=4.0, speed=1.0, seed=0, cpu_threads=8))
tts.synthesize("Bom dia!", speed=1.1, seed=7)   # per-call overrides

Command line

python -m rapido "O boleto vence no dia 05/11/2026." -o boleto.wav
python -m rapido -f texto.txt -o texto.wav

Portuguese, as written

Write normally. Rápido expands the written form into words before speaking, and tts.normalize(text) shows you exactly what it will say.

Benchmarks

Brazilian text is full of things written one way and spoken another: prices, dates, times, percentages, ordinals and phone numbers. We built a 40-sentence test of exactly that, gave every model the same raw text, transcribed the audio with Whisper large-v3, and compared the result word by word (numbers expanded to words before scoring).

Speed

Measured with plain PyTorch at 8 steps, one request at a time, taking the middle of 3 passes after a warm-up.

How it works

1. A convolutional text encoder reads Portuguese letters.
2. A duration predictor places each letter on a word–letter timeline.
3. A windowed Gaussian aligner gives every audio frame a soft mix of the letters near it.
4. A flow-matching diffusion transformer generates 64-dimensional latents at 25 Hz in 8 steps.
5. The frozen VoxCPM2 AudioVAE decodes the latents to 48 kHz.

Limitations

• One voice. There is no voice cloning and no emotion control.
• Portuguese only. Foreign names, English words and acronyms are read with Portuguese spelling rules.
• Narration style. The delivery is calm and read-aloud rather than conversational.
• Synthetic speech. Tell listeners when they are hearing an AI voice, and never use it to impersonate anyone.

Acknowledgements

The 48 kHz decoder is the AudioVAE from VoxCPM2 by OpenBMB (Apache 2.0).

License

Apache 2.0.

Citation

@misc{rapido2026,
  title        = {Rápido: Tiny, Fast Brazilian Portuguese Text-to-Speech},
  author       = {Kanade, Anurag},
  year         = {2026},
  howpublished = {\url{https://github.com/anurag12-webster/rapido-tts}}
}