Property	Value
Base model	nvidia/parakeet-tdt-0.6b-v3
Architecture	FastConformer-TDT (600M params)
Language	Portuguese (pt)
Input	16 kHz mono audio
Output	Portuguese text with punctuation and capitalization
License	CC-BY-4.0

Evaluation Results

Evaluated on Common Voice 17.0 Portuguese splits (raw text, no normalization):

Split	WER	CER	Samples
Validation	9.62%	--	9,464
Test	10.71%	2.69%	9,467

Training

Fine-tuned on the mixed_cv_synthetic configuration from yuriyvnv/synthetic_transcript_pt:

Common Voice 17.0 (pt) -- 21,968 human-recorded Portuguese speech samples
Synthetic samples -- 19,181 high-quality synthetic Portuguese speech samples generated with OpenAI TTS (filtered by audio-text similarity > 0.5)
Total training set: 41,149 samples

Validation and test sets use Common Voice 17.0 Portuguese exclusively (no synthetic contamination).

Training Configuration

Parameter	Value
Optimizer	AdamW
Learning rate	5e-5 (cosine annealing)
Warmup	10% of total steps
Batch size	64
Precision	bf16-mixed
Gradient clipping	1.0
Early stopping	10 epochs patience on val WER
Best epoch	15

Usage

Installation

pip install nemo_toolkit[asr]

Transcribe Audio

python

import nemo.collections.asr as nemo_asr

# Load model
asr_model = nemo_asr.models.ASRModel.from_pretrained(
    model_name="yuriyvnv/parakeet-tdt-0.6b-portuguese"
)

# Transcribe
output = asr_model.transcribe(["audio.wav"])
print(output[0].text)

Transcribe with Timestamps

python

output = asr_model.transcribe(["audio.wav"], timestamps=True)

for stamp in output[0].timestamp["segment"]:
    print(f"{stamp['start']:.1f}s - {stamp['end']:.1f}s : {stamp['segment']}")

Long-Form Audio

For audio longer than 24 minutes, enable local attention:

python

asr_model.change_attention_model(
    self_attention_model="rel_pos_local_attn",
    att_context_size=[256, 256],
)
output = asr_model.transcribe(["long_audio.wav"])

Intended Use

This model is designed for transcribing Portuguese speech to text. It works best on:

Read speech and conversational Portuguese
Audio recorded at 16 kHz or higher
Segments up to 24 minutes (or longer with local attention enabled)

Limitations

Trained on Common Voice data which is predominantly European Portuguese; performance may differ on Brazilian Portuguese
Synthetic training data was generated with OpenAI TTS voices, which may not fully represent natural speech variability
Not suitable for real-time streaming without additional configuration