Allosaurus isiZulu NCHLT — v0.1 Pilot

An experimental isiZulu phone-recognition model adapted from the Allosaurus uni2005 universal phone recognizer using NCHLT isiZulu speech.

This repository is a v0.1 pilot release. It is intended for research and reproducibility, not as a claim of solved or human-equivalent phonetic transcription for isiZulu.

Model description

The model was adapted from Allosaurus uni2005 to a 40-phone isiZulu-compatible target inventory.

Training targets were generated from the orthographic NCHLT isiZulu transcriptions using a validated Lwazi isiZulu grapheme-to-phoneme (G2P) pipeline and then mapped to an Allosaurus-compatible phone inventory.

The resulting model is therefore trained to recover canonical G2P-derived phone sequences from speech.

Important interpretation

The model has been evaluated against canonical G2P-derived phone sequences on held-out NCHLT speech.

It has not yet been comprehensively validated against human narrow phonetic transcriptions. Its outputs should therefore be interpreted as model-estimated phone sequences, not as ground-truth phonetic transcriptions of arbitrary pronunciation variation or pronunciation errors.

Model details

Item	Value
Base model	Allosaurus `uni2005`
Language	isiZulu (`zul`)
Release	v0.1 pilot
Training corpus	NCHLT isiZulu
Training utterances	4,313
Training duration	5.41 h
Validation utterances	1,019
Validation duration	1.23 h
Held-out test utterances	2,748
Held-out test duration	3.86 h
Held-out reference phones	66,676
Target inventory	40 phones
Best validation PER	14.189%
Adapted held-out test PER	13.59%
Zero-shot `uni2005` held-out test PER	44.85%

The pilot train/validation subsets were speaker-disjoint. Evaluation used the official held-out NCHLT test speakers.

Evaluation

Evaluation used corpus-level phone error rate (PER):

PER = (substitutions + deletions + insertions) / number of reference phones

Held-out NCHLT test results

Model	Substitutions	Deletions	Insertions	Total errors	Corpus PER
Allosaurus `uni2005` zero-shot	14,946	11,833	3,123	29,902	44.85%
isiZulu adapted v0.1	4,768	2,558	1,732	9,058	13.59%

The adapted model reduced PER by 69.71% relative to the zero-shot baseline on the held-out NCHLT test set.

The best validation PER was 14.189%, compared with 13.59% on the held-out test set.

The machine-readable results are provided in evaluation/model_comparison.csv.

Training data

The model was adapted using the NCHLT isiZulu speech corpus.

NCHLT audio is not redistributed in this repository.

The pilot used:

4,313 training utterances (~5.41 h)
1,019 validation utterances (~1.23 h)
2,748 official held-out test utterances (~3.86 h)

Utterances longer than 10 seconds were excluded during preparation for this pilot.

Because NCHLT contains prompted speech and may contain code-switched or non-isiZulu material, this release should not be interpreted as representing every isiZulu speech domain, accent, dialect, speaking style, or recording condition.

Phone targets

Orthographic NCHLT transcriptions were converted to canonical phone sequences using Lwazi isiZulu G2P rules. The Lwazi phone representation was then mapped to the target Allosaurus-compatible inventory.

This target construction is important when interpreting the reported PER: the reference sequences are canonical G2P-derived labels, rather than independent human narrow-phonetic annotations of each recording.

Intended use

This model may be useful for:

research on isiZulu phone recognition;
exploratory pronunciation analysis;
phone-level analysis of isiZulu speech;
comparison with zero-shot universal phone recognition;
low-resource speech technology experiments;
research prototypes in which the limitations of model-estimated phone sequences are explicitly acknowledged.

Out-of-scope / unsupported claims

This v0.1 model should not currently be treated as:

a ground-truth narrow phonetic transcriber;
a validated clinical or diagnostic pronunciation-assessment system;
a definitive detector of isiZulu mispronunciations;
evidence that every phone emitted by the model corresponds exactly to the phones acoustically produced by a speaker;
a model validated across all isiZulu accents, dialects, spontaneous-speech conditions, or demographic groups.

A future release should include evaluation against a human phone-annotated validation subset.

Repository structure

.
├── README.md
├── model/
│   ├── model.pt
│   ├── phone.txt
│   ├── am_config.json
│   ├── pm_config.json
│   ├── lm_config.json
│   └── inventory/
├── training/
│   └── training.log
├── evaluation/
│   └── model_comparison.csv
└── example/
    └── inference.py

Historical training checkpoints are intentionally not included in the release. model/model.pt is the deployable selected checkpoint.

Usage

This is a custom Allosaurus checkpoint, not a Transformers model. It is therefore not loaded with AutoModel.from_pretrained().

Install the required packages:

pip install allosaurus huggingface_hub

For a private repository, authenticate first:

hf auth login

Then edit REPO_ID in example/inference.py to use your Hugging Face namespace and run:

python example/inference.py path/to/audio.wav

The example downloads the Hugging Face snapshot, installs the packaged model into Allosaurus's local pretrained directory under the model name zul_nchlt_pilot_v3, and calls Allosaurus read_recognizer().

The example intentionally refuses to overwrite an existing local model directory unless --force-install is supplied.

Example:

python example/inference.py path/to/audio.wav --force-install

Reproducibility notes

This release preserves:

the selected model checkpoint;
acoustic/phone/language-model configuration files required by the packaged Allosaurus model;
the phone inventory resources packaged with the model;
the training log;
held-out evaluation totals.

The NCHLT audio itself is not included.

Limitations and future work

Planned improvements include:

training on a larger, carefully cleaned portion of NCHLT;
explicit handling and reporting of code-switched/non-isiZulu prompts;
evaluation across additional speaking conditions;
validation against human phone-transcribed isiZulu speech;
analysis of individual phone confusions and language-specific contrasts;
external evaluation beyond prompted NCHLT speech.

Licensing and attribution

License review is still required before making this repository public.

This model is derived from the Allosaurus software/model ecosystem and was adapted using NCHLT isiZulu data with labels generated using Lwazi isiZulu pronunciation resources. The applicable licenses and attribution requirements for Allosaurus, NCHLT, and Lwazi should be verified before assigning a final Hugging Face repository license.

The repository is therefore marked license: other for the initial private release rather than assigning an unverified permissive license.

Citation

If you use this pilot model in research, please cite the underlying Allosaurus, NCHLT, and Lwazi resources as appropriate.

A formal citation for this model/repository can be added once the public repository URL and release metadata are finalized.

Release status

v0.1-pilot

This release demonstrates a substantial in-domain improvement over zero-shot Allosaurus on held-out NCHLT isiZulu speech while retaining explicit limitations around the interpretation of G2P-derived phone references.

Downloads last month: -; Downloads are not tracked for this model. How to track