AccentScore 0.2.1 is a local, transcript-free English-pronunciation research model. It combines a frozen microsoft/wavlm-base-plus encoder with three task-specific heads:

a five-target pronunciation-quality regressor;
a four-class American/British/Indian/Russian accent classifier;
an American-vs-non-American classifier mapped to a 0–100 score.

The model accepts one 0.2–30 second audio recording and produces a structured PredictionResult. It does not call an external inference service.

Usage

Install AccentScore from its source repository, then load the pinned model bundle:

python

from accentscore import AccentScorer

scorer = AccentScorer.from_pretrained()
result = scorer.predict("recording.wav")
print(result.detected_accent, result.scores)

The model heads are downloaded once and cached locally. Inference remains local. The frozen WavLM encoder is also downloaded at the exact revision recorded in the bundle.

Intended use

Appropriate uses include local experimentation, pronunciation-practice interfaces, model evaluation, and research on representation transfer.

The model is not intended for employment, education access, immigration, identity inference, grading, diagnosis, or any decision that affects a person's rights or opportunities.

Accent labels refer only to resemblance among the supported acoustic classes. They are not claims about a speaker's nationality, ethnicity, birthplace, native language, or identity.

Architecture

Component	Configuration
Encoder	WavLM Base+, frozen, pinned revision
Input	Mono 16 kHz waveform
Pooling	Masked mean and population standard deviation from 13 layers
Quality head	Target-specific softmax layer mixtures and bounded MLP regressors
Accent head	Layer 7, L2 logistic regression, temperature scaling
American head	Layer 6, binary logistic regression, temperature scaling
Quality uncertainty	Split-conformal absolute residual intervals

The four-class and binary heads are intentionally independent. American similarity is not computed from pronunciation accuracy or from the four-class probability.

Training data

Component	Corpus	Split policy
Quality	SpeechOcean762 / OpenSLR SLR101	Official test preserved; train/validation/calibration are speaker-disjoint
Accent	AESRC2020 + Speech Accent Archive	Class-stratified, speaker-disjoint train/validation/test partitions
American similarity	AESRC2020	American versus British/Indian/Russian, speaker-disjoint

SpeechOcean762 is distributed under CC BY 4.0. Operators must independently verify AESRC2020 and Speech Accent Archive terms, consent, attribution, derivative-work, and redistribution requirements before using or distributing the accent checkpoints. The model weights are therefore marked license: other and are published for research and evaluation only; the source repository's MIT license applies to code, not to these weights or third-party datasets.

Evaluation

Pronunciation quality

SpeechOcean762 official test: 2,500 utterances from 125 speakers.

Metric	Macro result
Pearson	0.7487
Spearman	0.7486
MAE	6.109 points
RMSE	9.225 points
R²	0.4182
90% interval coverage	0.8922

Completeness is almost constant in this corpus. Its correlations are undefined and its near-perfect interval coverage should not be interpreted as strong modeling.

Accent tasks

Held-out source	Accuracy	Macro-F1
AESRC2020	0.9055	0.9047
Speech Accent Archive	0.9390	0.8950

The American-vs-non-American head reaches ROC-AUC 0.9607 on the speaker-disjoint AESRC test split.

The Hub bundle ships machine-readable results as model_evaluation.json; the source repository tracks the same receipt at artifacts/metrics/model_evaluation.json.

Output interpretation

american_similarity equals 100 times a calibrated binary probability. It is useful for ranking recordings under the model's training distribution, but it is not a direct percentage similarity and has not been calibrated against human accent-strength ratings.

The unknown label is emitted when no supported class reaches probability 0.4. This does not guarantee reliable out-of-distribution detection.

Pronunciation-quality intervals are marginal conformal intervals under exchangeability assumptions. They are not guarantees for an individual speaker or a shifted domain.

Known limitations

Four labels cannot represent the diversity or continuity of English pronunciation.
The training corpora may encode speaker, microphone, location, sentence, and source shortcuts.
Source-disjoint evaluation is limited to two accent corpora.
Short, noisy, code-switched, spontaneous, or unsupported speech can be unreliable.
Quality targets reflect one corpus's annotation conventions.
The model does not evaluate intelligibility, communication success, or personal ability.

Operational guidance

Keep the responsible-use warning with user-facing results.
Show probabilities and unknown, not only the top label.
Allow users to ignore or contest a result.
Validate on the target population and recording conditions before deployment.
Monitor performance separately by source and relevant acoustic conditions.
Do not infer protected or sensitive traits from accent outputs.

Reproducibility

The source repository pins dependencies in uv.lock, the model bundle is tagged v0.2.1, the encoder uses a full commit revision, and deterministic seeds and split rules live in the training code. Runtime consumers pin the full Hugging Face model commit rather than the mutable main branch.

Run static checks and the non-network test suite with:

make check

The exported model contract is covered by an opt-in real-encoder smoke test.

wiftwift/accent-scorerer

AccentScore model card

Summary