Instructions to use Eperson/eperson-voice with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use Eperson/eperson-voice with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-to-speech", model="Eperson/eperson-voice")# pip install -U transformers accelerate # Load model directly from transformers import AutoModelForSeq2SeqLM model = AutoModelForSeq2SeqLM.from_pretrained("Eperson/eperson-voice", device_map="auto") - Notebooks
- Google Colab
- Kaggle
eperson-voice: a Brazilian Portuguese (pt-BR) text-to-speech voice
eperson-voice is a single-speaker Brazilian Portuguese text-to-speech model with a natural, colloquial carioca narration style. It is Qwen3-TTS-12Hz-1.7B-Base fine-tuned on about 7.3 hours of the author's own voice, recorded for educational videos (philosophy, physics and math). It is the author's real voice, shared by the author.
| Language | Brazilian Portuguese (pt-BR) only |
| Speaker name | eperson |
| Style | Didactic, conversational narration (carioca accent) |
| Base model | Qwen3-TTS-12Hz-1.7B-Base (Apache 2.0) |
| Size | 1.7B parameters, about 4.5 GB (bf16) |
| Output | 24 kHz mono audio |
| License | Apache 2.0 |
Audio examples (sentences that were not in the training data): see the samples/ folder of this repository.
How to use
You need a GPU with about 6 GB of free VRAM (it also runs on CPU, but slowly) and Python 3.10+.
pip install -U qwen-tts soundfile torch
import torch
import soundfile as sf
from qwen_tts import Qwen3TTSModel
model = Qwen3TTSModel.from_pretrained(
"Eperson/eperson-voice",
device_map="cuda:0", # or "cpu"
dtype=torch.bfloat16,
attn_implementation="sdpa", # flash-attn is not required
)
torch.manual_seed(42) # fix the seed for a repeatable take
wavs, sr = model.generate_custom_voice(
text="Um vetor é, antes de tudo, uma lista de números em uma ordem.",
speaker="eperson",
language="Auto", # keep "Auto": the model was trained without a language tag
max_new_tokens=500, # safety cap (12.5 codec frames per second = about 40 s)
do_sample=True, temperature=0.8, top_k=30, top_p=0.85, repetition_penalty=1.05,
)
sf.write("out.wav", wavs[0], sr)
Tips for good results
- Write the text the way it is spoken, in pt-BR. Numbers and years are safer spelled out ("mil novecentos e cinco" instead of "1905").
- One sentence or short paragraph per call (up to about 15 seconds of speech), then join the audio files. For long narration, split the script into sentences, generate each one and concatenate with a short pause.
- Use
language="Auto". Passing"Portuguese"adds a language tag the model never saw during fine-tuning. - Re-generate with another seed if a take has an odd pause or a mispronounced word. It is normal to pick the best of a few takes.
Training details
- Data: 2,707 segments of 3 to 15 s (7.3 hours) cut at natural pauses from the author's narration, denoised with DeepFilterNet, 24 kHz mono. Transcripts come from Whisper word timing. A single fixed reference clip is used for all rows.
- Method: LoRA (rank 32, alpha 64, on all attention and MLP projections of the talker and the code predictor), learning rate 1e-4 with cosine decay, 2 epochs, effective batch size 8, bf16. The LoRA weights are merged, so this is a regular full checkpoint and no extra adapter is needed.
- Note on the training code: the official
sft_12hz.pyof Qwen3-TTS was not used because of known bugs (missingtext_projection, double label shift, code-predictor frame misalignment). A corrected training forward was used instead.
Limitations
- Only pt-BR, one speaker and one speaking style (explanatory narration). It does not do other languages, other voices or acted emotions.
- Occasional odd pauses mid-sentence, especially around commas.
- Rare proper names and foreign words can be mispronounced.
- No quality metrics beyond listening tests and an automatic transcription check; results can vary between seeds.
Responsible use
This model reproduces a real person's voice. Do not use it to impersonate the author or anyone else, to deceive listeners, or to put words in the author's mouth. Disclose synthetic audio where it could be mistaken for a recording.
License
Apache 2.0, the same as the base model. Built on Qwen3-TTS by the Qwen team (Alibaba Cloud).
- Downloads last month
- 211
Model tree for Eperson/eperson-voice
Base model
Qwen/Qwen3-TTS-12Hz-1.7B-Base