parakeet-tdt-0.6b-polish

Parakeet-TDT-0.6B Polish

A Polish automatic speech recognition (ASR) model fine-tuned from nvidia/parakeet-tdt-0.6b-v3.

Model Details

Evaluation Results

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

Training

Fine-tuned on a curated subset of the BIGOS v2 benchmark, filtered to retain only sources with proper casing and punctuation:

• Common Voice 15 -- 19,119 human-recorded Polish speech samples
• M-AILABS / LibriVox -- 11,834 read Polish audiobook samples
• PolyAI Minds14 -- 462 Polish banking dialog samples
• Total training set: ~31,415 samples

Validation uses the BIGOS v2 validation split (same source filtering). Test evaluation uses Common Voice 17.0 Polish (independent test set).

Training Configuration

Usage

Installation

pip install nemo_toolkit[asr]

Transcribe Audio

import nemo.collections.asr as nemo_asr

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

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

Transcribe with Timestamps

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:

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 Polish speech to text. It works best on:

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

Limitations

• Training data is sourced from read speech (audiobooks, Common Voice read prompts) and short banking dialogs; performance may differ on spontaneous or heavily accented speech
• The model preserves punctuation and capitalization as seen in training data
• Not suitable for real-time streaming without additional configuration