parakeet-tdt-0.6b-an4

This is a fine-tuned version of NVIDIA's nvidia/parakeet-tdt-0.6b-v3 automatic speech recognition (ASR) model, adapted to the AN4 dataset using the NVIDIA NeMo framework.

The base model is a 600M-parameter FastConformer model with a Token-and-Duration Transducer (TDT) decoder. This fine-tune adapts it to AN4's specific output style — lowercase text, no punctuation, and spelled-out letters and digit strings.

Model description

• Base model: nvidia/parakeet-tdt-0.6b-v3 (FastConformer encoder + TDT decoder)
• Architecture: EncDecRNNTBPEModel (RNN-Transducer with BPE tokenization)
• Fine-tuned on: AN4 (English, spelled names / addresses / digit strings)
• Language: English
• Framework: NVIDIA NeMo

Intended use and limitations

Intended use. This model is a learning/demonstration artifact showing ASR fine-tuning on a small, domain-specific dataset. It transcribes short English utterances in AN4's style (spelled letters, spoken digits, short commands).

Important limitations — read before using:

• Highly specialized. It was fine-tuned only on AN4, which contains spelled names, addresses, dates, and digit strings — not natural sentences. It performs well on AN4-style speech and will behave poorly on general conversational or read speech.
• Output style. It emits lowercase text with no punctuation or capitalization, because that is AN4's label format. It does not restore casing or punctuation like the base model does.
• Not production-grade. AN4 is a tiny dataset (948 training utterances). This model is for education, not deployment.
• For general English ASR, use the base model nvidia/parakeet-tdt-0.6b-v3 instead.

How to use

Install NeMo, then load and transcribe with the NeMo toolkit:

```python import nemo.collections.asr as nemo_asr

model = nemo_asr.models.ASRModel.from_pretrained("bhaveshsoni0023/parakeet-tdt-0.6b-an4")

Audio must be 16 kHz mono WAV/FLAC

hyps = model.transcribe(["path/to/audio.wav"]) for h in hyps: print(h.text if hasattr(h, "text") else h) ```

Training details

Fine-tuned with the `speech_to_text_finetune.py` script from the NeMo Speech repository, warm-started from the pretrained base model (the base tokenizer was kept, i.e. no tokenizer update).

Setting	Value
Base model	`nvidia/parakeet-tdt-0.6b-v3`
Dataset	AN4 (train: 948 utts, test: 130 utts)
Tokenizer	Reused from base model (no change)
Optimizer	AdamW
Learning rate	1e-4
LR schedule	CosineAnnealing, 100 warmup steps
Batch size	16
Epochs	20 (1200 steps)
Precision	bf16
Loss	TDT (Token-and-Duration Transducer)
Sample rate	16 kHz mono

Best validation WER during training reached ~0.041 (4.1%).

Evaluation

Evaluated on the AN4 test split (130 utterances), Word Error Rate (WER, lower is better):

Model	Test WER
Baseline (`nvidia/parakeet-tdt-0.6b-v3`)	75.81%
Fine-tuned (this model)	0.39%

How to read these numbers (important). The large gap is partly real adaptation and partly a scoring artifact. The baseline model outputs punctuation and capitalization and expects natural sentences, so on AN4's bare lowercase spelled-out labels it is penalized heavily by WER even when it recognizes the speech correctly. The fine-tuned model scores far lower mostly because it learned to match AN4's exact output format. A fairer comparison would normalize both hypotheses and references (lowercase, strip punctuation) before scoring, which would raise the baseline substantially and shrink the true gap. The fine-tune does genuinely help, but the headline delta overstates it.

License and attribution

This model is released under CC-BY-4.0, inherited from the base model.

It is a derivative of `nvidia/parakeet-tdt-0.6b-v3` by NVIDIA, which is licensed under CC-BY-4.0. Attribution to NVIDIA is required under the terms of that license. The AN4 dataset is used for training/evaluation.

Citation

If you use this model, please also credit the base model and framework:

• Base model: NVIDIA Parakeet TDT 0.6B v3 — https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3
• Framework: NVIDIA NeMo — https://github.com/NVIDIA-NeMo/Speech