manassehzw/sna-parakeet-tdt-0.6b-v3

Shona Parakeet TDT 0.6B

This is a Shona (sna) automatic speech-recognition model fine-tuned from `nvidia/parakeet-tdt-0.6b-v3` using human-labelled Shona speech.

The model is fast and produces punctuation and capitalization. Its strongest result is on WAXAL-style speech, while performance drops on FLEURS and biblical narration. It is useful as a fast research recognizer and adaptation baseline, but it is not a general-purpose production Shona model.

Model details

Evaluation

The original card contained placeholder metrics. The results below come from saved predictions produced by model revision f588f40745b4944491feb0b9a3eb3ede6663b6c3.

Standard project scorer

The standard scorer applies Unicode NFKC, case-folding, and whitespace collapse. It retains punctuation. Percentages are corpus-level WER.

Punctuation-insensitive rescore

This scorer additionally replaces Unicode punctuation with spaces before calculating WER. It separates lexical recognition from punctuation style.

^* This model was trained with an older random 90/5/5 split over the annotated WAXAL corpus. That split does not match the later published WAXAL partition, so overlap with the current WAXAL test cannot be ruled out. Treat the WAXAL scores as contextual rather than strictly held out. FLEURS and the curated recordings do not have that fine-tuning split-overlap issue.

FLEURS is public and may have appeared in foundation-model training data. The curated set is temporally separate from model pretraining, but it contains one speaker and only 14.9 minutes of clean audio. Neither benchmark fully represents noisy, multi-speaker, real-world use.

Inference profile

In the shared NVIDIA L4 benchmark, this checkpoint processed audio at approximately 206.4 times real time with batch size 8 and used 4.05 GiB peak VRAM. These figures exclude model loading and depend on hardware, batching, audio length, and software versions.

Intended use

Appropriate uses include:

• —fast Shona transcription research on segmented speech;
• —a baseline for further Shona or domain-specific adaptation;
• —timestamp and alignment experiments using NeMo decoding outputs;
• —comparison with multilingual and teacher ASR models.

Evaluate on data that matches the intended domain before using automatically generated transcripts for training or publication.

Limitations

• —Accuracy drops sharply outside WAXAL-style speech, particularly on FLEURS and biblical narration.
• —The model has not been validated as a streaming recognizer.
• —Long recordings should be segmented before transcription.
• —Code-switching, telephone audio, regional variation, noise, and overlapping speakers have not been evaluated systematically.
• —Punctuation and capitalization are model predictions and may differ from reference style.
• —The WAXAL score has the historical split caveat described above.

Usage

The repository stores a NeMo .nemo artifact. Download it and restore it explicitly:

from huggingface_hub import hf_hub_download
import nemo.collections.asr as nemo_asr

repo_id = "manassehzw/sna-parakeet-tdt-0.6b-v3"
model_path = hf_hub_download(repo_id, filename="sna-parakeet-v1.nemo")
model = nemo_asr.models.ASRModel.restore_from(model_path)
model.eval()

results = model.transcribe(["audio-16khz-mono.wav"])
transcript = results[0].text if hasattr(results[0], "text") else str(results[0])
print(transcript)

Use 16 kHz mono audio. Move the model to the desired device before inference, and segment long recordings while preserving source boundaries.

Training provenance

The historical preprocessing summary is retained in training/preprocess_summary.json. The run used 13,715 training examples, 762 validation examples, and 762 test examples from a 15,239-example corpus, with seed 42. This was an early local split and differs from the dataset's later published 12,170/1,504/1,565 partition.

The exact optimizer and scheduler recipe was not preserved in this model repository. The card therefore reports the facts supported by the uploaded artifacts and saved evaluation predictions rather than filling missing fields with guesses.

Repository files

• —sna-parakeet-v1.nemo: exported NeMo model artifact;
• —training/preprocess_summary.json: historical data-preparation record;
• —eval/last_inference.json: three qualitative transcription examples.

License and attribution

This fine-tune is released under CC-BY-4.0, matching the `nvidia/parakeet-tdt-0.6b-v3` base checkpoint and the annotated Shona training dataset. Review the upstream model and dataset cards for their attribution requirements.

Relevant sources:

• —NVIDIA Parakeet TDT 0.6B v3
• —Annotated Shona fine-tuning dataset
• —WAXAL NLP
• —FLEURS

Citation

If you use this Shona fine-tune, cite this repository:

@misc{changachirere2026shonaparakeet,
  author       = {Manasseh Changachirere},
  title        = {Shona Parakeet TDT 0.6B},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/manassehzw/sna-parakeet-tdt-0.6b-v3}}
}

Please also cite NVIDIA's base model and the source dataset when required by your use.

Contact

Open a discussion in this Hugging Face repository for questions, reproducibility notes, or evaluation corrections.