HRB25/inflect-micro-ptbr-20k

Inflect Micro PT-BR · 20k

Inflect Micro PT-BR is a single-speaker Brazilian Portuguese TTS model adapted from the open-weight owensong/Inflect-Micro-v2. It was trained on 20,000 clips (37.45 hours) from one speaker at 24 kHz, for 20,000 steps, and is released as a self-contained PyTorch inference package.

Listen to the samples ↓ · Get started ↓ · See the training evidence ↓ · Read the training recipe ↓

Table of contents

• Model Overview
• Results at a Glance
• Listen
• Quick Start
• Model Architecture
• Inputs and Outputs
• Training
• Evaluation
• Engineering Notes
• Limitations
• Reproduction
• License
• Citation

Model Overview

Description

Inflect Micro PT-BR adapts a multilingual VITS-based speech model to one Brazilian Portuguese speaker. The adaptation uses a staged schedule: the posterior encoder warms up first, the linguistic layers follow, and the decoder stays frozen until step 6,000 so the acoustic quality learned from the base model survives the language shift.

The package in this repository runs on CPU or GPU without installing the training toolchain. Text goes in, a 24 kHz waveform comes out, with adjustable speed and prosody variation.

This model is a research artifact. Read Limitations before deploying it.

Model Summary

Base model	owensong/Inflect-Micro-v2
Architecture	VITS (posterior encoder, flow, HiFi-GAN decoder), inference-only export
Parameters on disk	37.5 MB (model.pth)
Sample rate	24 kHz, mono, PCM
Language	Brazilian Portuguese (pt-BR)
Speakers	1 (source speaker ID 2961)
Text frontend	espeak-ng, pt-br, with stress marks and punctuation preserved
Symbol inventory	181 entries (178 from the base model + 3 added for Portuguese)
Inputs / Outputs	Text / 24 kHz waveform
Control knobs	--speed (duration scaling), --variation (prosody diversity), --seed
Training hardware	1× A100 40 GB, ~3.2 GB peak VRAM
Training budget	20,000 steps, effective batch 8
License	Inherited from the base model; see License

Results at a Glance

Training losses are means over step ranges, computed from 22,404 rows of metrics.jsonl. ↓ Lower is better.

Step range	loss_g	loss_d	What was happening
1–999	75.22	2.192	Posterior warm-up, decoder frozen
1,000–4,999	45.84	1.857	Discriminator learns the base/target gap
5,000–9,999	29.97	2.562	Decoder unfreezes at 6,000
10,000–14,999	26.13	2.695	Timbre consolidation
15,000–19,499	25.64	2.687	Plateau
19,500–20,000	25.53	2.679	Final checkpoint

loss_g drops 66% across the run. loss_d falls early, then rises and settles near 2.68 — the expected rebalancing once the generator starts producing the target speaker's distribution and the discriminator loses its shortcut.

Listen

Every clip below is served from this repository. Press play; nothing is downloaded separately.

Prosody control

Same eight sentences, same seed (42), CPU inference. Only speed and variation change, so the prosody is the single variable.

ID	Sentence	default speed 1.0 · var 0.667	slower speed 0.90 · var 0.667	varied speed 0.95 · var 0.80
frase-00	Olá, tudo bem com você?	1.65 s	1.78 s	1.72 s
frase-01	Hoje está fazendo um dia muito bonito.	2.55 s	2.73 s	2.61 s
frase-02	O Brasil é um país de dimensões continentais.	3.43 s	3.63 s	3.53 s
frase-03	Você gostaria de ouvir uma história interessante?	3.11 s	3.33 s	3.23 s
frase-04	Não acredito que isso aconteceu!	2.20 s	2.25 s	2.23 s
frase-05	Que notícia maravilhosa!	1.76 s	1.90 s	1.83 s
frase-06	Por favor, me diga que horas são.	2.10 s	2.29 s	2.19 s
frase-07	Amanhã teremos uma reunião importante.	2.60 s	2.76 s	2.65 s

Total duration: 19.41 s (default), 20.67 s (slower), 19.99 s (varied). Reducing speed to 0.90 stretches the mean utterance by 6.6%, not the 11.1% a purely linear scaling would give — the model targets per-phoneme duration, and pauses do not scale with it. Short sentences deviate most: frase-04 stretched 2.4%, while frase-02 stretched 9.1%.

Version comparison: 4k steps vs 20k steps

Same six sentences, synthesized by the earlier 4,000-step checkpoint and by this 20,000-step one. Meant for listening, not for crowning a winner.

Sentence	V1 · 4k steps	V2 · 20k steps
Bom dia! Tudo bem com você?	V1 · 2.34 s	V2 · 2.23 s
O Brasil é um país de dimensões continentais.	V1 · 3.13 s	V2 · 3.51 s
Preciso marcar uma consulta para a próxima terça-feira.	V1 · 3.66 s	V2 · 3.67 s
A reunião começou às nove da manhã e terminou ao meio-dia.	V1 · 3.59 s	V2 · 3.63 s
Eu gostaria de um café com leite e dois pães de queijo, por favor.	V1 · 4.27 s	V2 · 4.31 s
O relatório será entregue até o final da semana que vem.	V1 · 3.62 s	V2 · 3.62 s

Training progression

The model's own validation samples, taken every 500 steps. Useful for hearing where the voice becomes intelligible and what changes after the decoder unfreezes.

Step	Stage	Validation audio
500	Primeiros passos: posterior aquecendo, decoder congelado.
5.000	Posterior estabilizado, ainda sem descongelar o decoder.
10.000	Decoder descongelado em 6.000; ajuste de timbre.
15.000	Fase final de refinamento.
20.000	Checkpoint final exportado.

Quick Start

The package is self-contained. You do not need inflect-adapt to run inference.

pip install -r requirements.txt

python inference.py --text "Bom dia, tudo bem com você?" --output saida.wav

With prosody control:

python inference.py \
  --text "A reunião começou às nove da manhã." \
  --speed 0.9 \
  --variation 0.8 \
  --device cuda \
  --output saida.wav

Batching over a jsonl manifest:

while IFS= read -r line; do
  id=$(echo "$line"  | python3 -c "import sys,json;print(json.load(sys.stdin)['id'])")
  txt=$(echo "$line" | python3 -c "import sys,json;print(json.load(sys.stdin)['text'])")
  python inference.py --text "$txt" --output "audios/$id.wav"
done < frases.jsonl

Flag	Default	Effect
--text	—	Plain text input, normalized by the espeak frontend
--phonemes	—	Skip the frontend and pass phonemes directly
--output	required	Destination .wav
--device	cpu	cpu or cuda
--speed	1.0	Duration multiplier across the utterance
--variation	0.667	Prosody diversity; higher is looser
--seed	0	Reproducibility seed

Model Architecture

The export is a VITS-style end-to-end TTS model: a text encoder, a normalizing flow, a posterior encoder used at training time, and a HiFi-GAN decoder that upsamples to 24 kHz.

Component	Configuration
Inter channels / hidden channels	192 / 96
Encoder layers, heads	3, 2
Upsample rates	8, 8, 2, 2
Upsample kernel sizes	16, 16, 4, 4
Resblock kernel sizes	3, 7, 11
Mel channels, FFT	80, 1024
Hop length	256

The Portuguese adaptation added three symbols — õ, ũ and the combining tilde ̃ — to the base inventory. Before adaptation, 97.6% of corpus symbols already existed in the base vocabulary, so the embedding table grew by three rows while staying compatible with the pretrained weights.

Inputs and Outputs

Property	Specification
Input types	Plain text, or phoneme sequences with --phonemes
Output type	Mono waveform, float32
Output format	24 kHz .wav
Frontend	espeak-ng pt-br, stress marks on, punctuation preserved
Text limit	The package splits long input into chunks at punctuation boundaries
Determinism	Fixed --seed and --variation 0 reproduce the same audio

Chunking is worth knowing about: sentences longer than the internal limit are synthesized in pieces and concatenated, which can produce a prosody discontinuity at the junction.

Training

Dataset

Metric	Value
Total clips	20,000
Train / validation	19,123 / 877
Total duration	134,807 s (37.45 h)
Clip duration min / max	1.50 s / 12.00 s
Sample rate	24 kHz
Speakers	1
Language / frontend	pt-BR / espeak with stress

The split is group-aware and normalizes text before separating, so the same sentence never appears in both sides. The target was 1,000 validation rows; the result was 877 because whole groups were kept together rather than sliced. Smaller validation set, no text leakage between train and evaluation.

Recipe

Parameter	Value
Base checkpoint	owensong/Inflect-Micro-v2
Preset	micro-12gb
Steps	20,000
Batch × accumulation	4 × 2 (effective 8)
Learning rate (G / D)	8e-5 / 8e-5
Posterior warm-up	1,500 steps
Decoder unfreeze	step 6,000
Precision	AMP
Checkpoint / validation interval	500 steps
Seeds (training / split)	1234 / 42

Final learning rates were 6.95e-5 for the posterior, 3.48e-5 for the linguistic layers and 6.95e-6 for the decoder. The asymmetry is deliberate: the decoder moves slowly so the acoustic detail from the base model is not overwritten.

Evaluation

The prosody runs used inflect-adapt evaluate on CPU with a fixed seed. All three configurations passed the automatic checks.

Configuration	Evaluated	Failures	Clipped clips	Silent clips	Result
default	8/8	0	0	0	PASS
slower	8/8	0	0	0	PASS
varied	8/8	0	0	0	PASS

These checks verify that the audio is well-formed and audible. They do not measure intelligibility. See Limitations.

Engineering Notes

Three problems cost real time during this run. All three share a cause: state that survives a restart and nobody reconciles.

Concurrent training processes. Restarting the run after a session drop spawned a new process without killing the old one. Within a few cycles, five inflect-adapt processes were sharing one A100 and writing to the same latest.pth. The symptom was not a clean error — the loss moved backwards, because different copies saved checkpoints with different optimizer states. Restart routines now terminate the process tree by PID and abort if anything survives.

Non-monotonic W&B steps. A resumed run emits step numbers below the ones already logged. W&B rejects non-increasing steps, so every new row was discarded while training continued normally. The fix separates the two axes: the W&B step counter stays monotonic, and the real training step is logged as a metric.

Validation audio that seemed to stop. The uploader checked for new metric lines first and returned early when there were none — before looking at audio. Since metrics are written every 25 steps and validation every 500, audio only uploaded when the two coincided. Metrics and audio are now independent in the same cycle, and the list of already-sent clips is persisted to disk so restarts do not resend old samples.

Limitations

• No intelligibility metric. The transcript evaluator was disabled in every run. "The voice is intelligible" remains an auditory judgment, not a measured result.
• No MOS study. No listening test was conducted.
• Single speaker. One voice. The model does not accept speaker_id and does not generalize timbre.
• Narrow domain. Read speech, 1.5–12 s per clip. Long utterances are chunked, which can cause prosody seams.
• Text normalization gaps. Acronyms, spelled-out numbers and foreign terms depend on the espeak frontend and were not systematically evaluated.
• License. See below. Confirm the base model and corpus terms before commercial use.

Reproduction

Path	Contents
training/launch-config.yaml	Effective configuration recorded by the run
training/train.py	Launcher with resume and W&B synchronization
training/supervisor.py	Automatic restart from the newest intact checkpoint
training/hyperparameters.yaml	Hyperparameters
validation/	40 audio and metadata pairs captured every 500 steps
audio/prosody_tests/	Prosody test across three configurations
checksums.sha256	Checksums for the exported package
export_report.json	Export verification report

Training ran at 1.25–2.0 s per step on one A100 40 GB. The source corpus is not redistributed here for licensing reasons; launch-config.yaml records the paths that were used.

License

The package carries the license of the base model. Review the terms of owensong/Inflect-Micro-v2 and of the source corpus before redistribution or commercial use.

Citation

@misc{rios2026inflectptbr,
  author = {Hugo Rios},
  title  = {Inflect Micro PT-BR: 20k steps of single-speaker adaptation},
  year   = {2026},
  url    = {https://huggingface.co/HRB25/inflect-micro-ptbr-20k}
}

Built on Inflect by owenawsong.