tiny-broca

tiny-broca is a small Brazilian Portuguese text-to-speech model with two voices, F1 (female) and M1 (male). The complete model, including the waveform generator, has 4.5M parameters and was trained from scratch on 17 hours of synthetic speech.

Parameters 4,533,835
Voices F1, M1 (Brazilian Portuguese)
Output 24 kHz mono
Input Text, tokenised as letters (58 symbols), no dictionary
Pretrained weights in the model None
License Apache-2.0

Files

File Contents
tiny-broca.pt PyTorch weights, float32
onnx/model.onnx ONNX export, float32, opset 17
config.json Released settings: the model block (sample rate, windows, voices, vocabulary size) plus the phonemiser identity, letter vocabulary, and ONNX inputs/outputs that the web demo reads
text.py The front end module, verbatim (the web demo runs it in the browser)
samples/ The audio on this page

Samples

Text F1 M1
A reunião de amanhã foi remarcada para as quinze horas e quinze.
Ligue para a sua avó antes que comece o fim de semana, por favor.
Quantas pessoas vão à conferência em setembro?
Não acredito que já é meia-noite e meia; devemos ir embora agora!
A tiny broca é um pequeno modelo de texto para fala treinado do zero.
Depois da tempestade, o ar da montanha parecia fresco, limpo e surpreendentemente quente.
Você pode me enviar o relatório até sexta-feira à tarde, por favor?
A conta do aluguel fechou em mil duzentos e quarenta e cinco reais e dez centavos.

Usage

PyTorch

pip install git+https://github.com/brenorb/tiny-broca
hf download brenorb/tiny-broca --local-dir tiny-broca

python -m tiny_broca.infer \
  --checkpoint tiny-broca/tiny-broca.pt \
  --voice F1 \
  --text "Olá! Eu sou a tiny broca." \
  --output ola.wav
Option Default Meaning
--voice F1 F1 or M1
--speed 1.0 Speaking rate
--pitch-shift 0 Added to log F0. 0.058 is about one semitone.
--pitch-range 1.0 Scales pitch movement around the utterance mean.

ONNX

The graph takes letter ids and stress values and returns the waveform. Text is converted to ids outside the graph by the front end (text.py); the ids are the positions in its letter inventory.

import numpy as np
import onnxruntime as ort
import soundfile as sf
import sys
sys.path.insert(0, "tiny-broca")
from text import phonemize, SYMBOL_TO_ID

letters = phonemize("Olá! Eu sou a tiny broca.")
one = lambda value: np.array([value], dtype=np.float32)

session = ort.InferenceSession("tiny-broca/onnx/model.onnx")
audio, durations = session.run(None, {
    "tokens": np.array([letters.ids], dtype=np.int64),
    "stress": np.array([letters.stress], dtype=np.int64),
    "voice": np.array([0], dtype=np.int64),   # 0 = F1, 1 = M1
    "speed": one(1.0),
    "pitch_shift": one(0.0),
    "pitch_range": one(1.0),
})
sf.write("ola.wav", audio[0], 24000)
Input Type Shape Meaning
tokens int64 [1, N] Letter ids, including the start and end tokens
stress int64 [1, N] 0 unstressed, 1 stressed syllable
voice int64 [1] 0 = F1, 1 = M1
speed float32 [1] 1.0 as predicted; larger is faster
pitch_shift float32 [1] Added to log F0
pitch_range float32 [1] Scales pitch movement
Output Type Shape Meaning
audio float32 [1, samples] Waveform at 24 kHz
durations int64 [1, N] Length of each token in frames of 256 samples

The model reads one utterance per call, up to 512 tokens and 32 seconds. Split longer text into sentences. The same file runs in onnxruntime-web (WebAssembly); the page at the demo URL speaks in the browser.

Model

Text is tokenised as letters (the 26 letters plus the accented graphemes that are distinct letters in Portuguese), with a stress value per letter. A transformer encoder (4 blocks) reads the letters, a duration predictor gives each one a length in frames, and two frame-level transformers (2 and 5 blocks) predict pitch, voicing and energy and then the parameters of the waveform generator. All blocks use 192 channels, 4 heads and rotary position embeddings.

The waveform is produced by a source-filter synthesizer instead of a neural vocoder. A band-limited pulse train at the predicted pitch and a noise signal are each shaped by a time-varying filter, whose cepstrum the network predicts once per frame (256 samples, 10.7 ms). This part has 25,862 parameters.

Component Parameters
Embeddings (letter, stress, voice) 13,248
Text encoder 1,588,864
Duration and pause predictor 123,906
Frame transformer A, prosody head 795,971
Frame transformer B 1,985,984
Source-filter waveform head 25,862
Total 4,533,835

Training

Data brenorb/tiny-broca: 9,384 sentences, 16,434 clips, two voices
Steps 118,000
Losses Multi-resolution spectral losses; adversarial loss with multi-period and multi-resolution discriminators after step 8,000
Hardware Apple M4 Pro, 18-core, MPS

The training audio is synthetic: every clip was generated with Qwen3-TTS 1.7B Base (Apache-2.0) in one of the two voices. The voices are clones of two Common Voice speakers (CC0): short clean recordings of real people, used only as voice prompts for the generator. No recording of a real speaker is part of the model. Letter durations come from forced alignment with jonatasgrosman/wav2vec2-xls-r-1b-portuguese, which is used only to label the training data. No pretrained weights are part of the model.

The first 104,000 steps use batch1 and batch2. The released model then continues for 14,000 more steps (to step 118,000) on the same data plus batch3, money phrases: the generator tends to rush long number strings, so the later rounds give money values extra weight.

Evaluation

Word error rate of synthesized speech, transcribed with Whisper (multilingual, pt). The recordings of the same sentences are scored the same way for reference. (Pending the training run.)

Limitations

  • Brazilian Portuguese only, two fixed voices, no voice cloning.
  • The voice is clear but can sound slightly synthetic.
  • The same text is always read the same way. There is no sampling of prosody.
  • Very long number strings (a dozen or more words) are read a little faster than the training data; the duration predictor compresses them slightly.
  • Foreign names and unusual spellings are often mispronounced.
  • No streaming.
  • The training data is itself synthetic, so the model cannot be more natural than the model that generated it.

License

The weights and code are released under Apache-2.0. No pronouncing dictionary is used anywhere: the front end is self-contained. The Common Voice reference recordings are CC0.

Citation

@software{brenorb2026tinybroca,
  author = {brenorb},
  title  = {tiny-broca: a 4.5M-parameter two-voice Brazilian Portuguese text-to-speech model},
  year   = {2026},
  url    = {https://huggingface.co/brenorb/tiny-broca}
}

Acknowledgements

The project replicates the open English two-voice model by Muhammed Nazeem (tiny-voca, Apache-2.0): same architecture, training recipe and release methodology, adapted to Brazilian Portuguese. The design builds on FastSpeech 2 (explicit duration and prosody prediction), DDSP and the neural homomorphic vocoder (source-filter synthesis with predicted filters), and HiFi-GAN and UnivNet (discriminators). No code or weights from these projects are used.

Downloads last month
5
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Dataset used to train brenorb/tiny-broca