FlexAligner Models

Official acoustic-model bundles and reference pronunciation resources

GitHub · PyPI · English · 简体中文

This repository hosts the model assets used by USTCPhonetics/FlexAligner. Use an immutable tag or commit SHA for reproducible work; main is the candidate asset branch for the next package release.

English

Language	Assets	Status
English	en/chunker, en/aligner	Published and release-tested by FlexAligner
Mandarin	zh/chunker, zh/aligner, zh/word.dict	Updated AISHELL-1 candidate for the next FlexAligner release

The English model files are unchanged from the release-tested bundle. English pronunciations use a user lexicon with the package's optional local G2P fallback; this repository does not ship an English system dictionary.

The Mandarin bundle contains tone-bearing rimes, with 208 Chunker tokens and 207 Aligner tokens. The Aligner supports both sil and sph. The supplied zh/word.dict is an AISHELL-1 reference lexicon. The current model-compatible pronunciation for ri uses r iz4, for example 日 r iz4.

Compatibility boundary

• FlexAligner 0.3.0a1 remains pinned to the older immutable model commit and is not compatible with the updated Mandarin files on main.
• The next FlexAligner version must add tone-preserving Chunker lexicon mapping, tone-bearing Mandarin G2P, Mandarin sph, updated language detection, and a new exact manifest pin before using this bundle automatically.
• Existing immutable revisions and tag v0.2.0a1 are retained unchanged.

Repository layout

.
├── README.md
├── model_manifest.json
├── en
│   ├── chunker
│   └── aligner
└── zh
    ├── chunker
    ├── aligner
    └── word.dict

model_manifest.json records the exact file set, sizes, SHA-256 digests, compatibility boundary, and provenance statements.

Training-data provenance

• The English models were trained on the LibriSpeech ASR corpus (OpenSLR SLR12), whose official resource page lists the dataset license as CC BY 4.0.
• The Mandarin models were trained on AISHELL-1 (OpenSLR SLR33), whose official resource page lists the dataset license as Apache License 2.0.
• Exact training subsets, recipes, seeds, and trainer identity are not included in the supplied model bundle. No unverified accuracy or domain-coverage claim is made.

License and provenance

USTCPhonetics publishes and redistributes the model weights with authorization to make these artifacts public. It does not claim ownership of the underlying LibriSpeech or AISHELL-1 datasets.

The Mandarin dictionary is the pre-supplied reference lexicon corresponding to AISHELL-1's supplementary resources. It is provided for reference under the AISHELL-1 resource's Apache-2.0 license; USTCPhonetics, OpenPhonetics, and FlexAligner make no ownership claim over that lexicon.

The repository uses license: other because it contains assets with different provenance and license layers. The dataset licenses do not silently replace the terms applicable to model weights, and the MIT license of the FlexAligner source repository is not represented as ownership of the datasets or lexicon.

简体中文

本仓库保存 FlexAligner 使用的中英文模型资产。为保证可复现性，请固定不可变 tag 或完整 commit SHA；main 是下一软件版本的候选模型分支。

• 英语模型文件保持不变，训练数据为 LibriSpeech SLR12（官方页面标注 CC BY 4.0）。
• 新普通话模型训练数据为 AISHELL-1 SLR33（官方页面标注 Apache-2.0）。
• 普通话 Chunker/Aligner 词表分别为 208/207，韵母保留 1--5 调号，Aligner 支持 sil 和 sph。
• zh/word.dict 是预先提供的 AISHELL-1 对应参考词典；本项目不声明对该词典的所有权。
• 当前规则使用 ri -> r iz4，例如 日 r iz4。
• 已发布的 FlexAligner 0.3.0a1 仍固定旧模型 revision，不能直接使用 main 上的新 普通话 bundle；下一软件版本完成兼容改造和真实模型 E2E 后才能切换。

本仓库采用 license: other，用于准确表达模型权重、训练数据和参考词典的分层来源； FlexAligner 源代码的 MIT 许可不等于对底层数据或词典主张所有权。