Want to create a speech dataset, but current forced alignment tools don’t work for your setup?

Mana Forced Aligner is a robust, language-agnostic forced alignment module designed specifically for low-resource languages and imperfect audio-text matches. Unlike traditional tools like Aeneas, our method does not require a perfect match between audio and text. It tolerates skipped words, repetitions, or inconsistencies, using multiple ASR models and character-level text similarity with predefined thresholds.

🧠 ASR-Agnostic and Scalable: Use one or more automatic speech recognition (ASR) models, regardless of their quality. The aligner is designed to work even with imperfect ASRs. However, the more ASR models you provide, the more robust and accurate the alignment becomes, thanks to majority-voting and fallback mechanisms.

🧩 Mismatch-Tolerant: Handles skipped or added phrases, word reorderings, and slight transcription errors.

🧪 Flexible Scoring: Matches based on character error rate (CER), with configurable thresholds.

🔓 Open and Extendable: Built in Python, licensed under MIT, and ready to adapt to your language and ASR tools.

🌍 Proven in Practice: Successfully used to generate 102+ hours of aligned speech-text data for the Persian ManaTTS corpus.

The alignment process consists of two key components:

• Runs multiple ASR models on each audio chunk
• Discards unreliable transcripts (e.g., truncated outputs)
• Sorts by reliability and returns top transcripts

Figure: Architecture of the transcription module, showing multiple ASRs and reliability filtering.

• Splits audio into 2–12 second segments based on silence

• Uses the transcription module to hypothesize possible texts

• Finds best-matching substrings in the reference text

• Accepts matches based on CER:

  • High quality: CER ≤ 0.05
  • Middle quality: 0.05 < CER ≤ 0.2

• Chunks with no acceptable match are discarded

The algorithm uses Interval Search first, then Gapped Search only if needed.

Figure: Forced alignment pipeline showing silence splitting, transcription-based matching, CER thresholds, and chunk decisions.

Dataset Name	Language	Size	License	Links
ManaTTS	Persian	102+ hrs	CC-0
Quran-Persian	Persian	20+ hrs	CC-0

Feel free to reach out if you'd like yours featured.

You can use the forced alignment method via Colab or the forced alignment notebook in this repository.

💡 To adapt it for your language, simply embed one or more ASR models (the more the better!) that work for your language into the transcription module section of the notebook.

• Actively tested on Persian
• Easily customizable for other low-resource languages with available ASR models

If you use Mana Forced Aligner in your work, please cite:

@inproceedings{qharabagh-etal-2025-manatts,
    title = "{M}ana{TTS} {P}ersian: a recipe for creating {TTS} datasets for lower resource languages",
    author = "Qharabagh, Mahta Fetrat  and Dehghanian, Zahra  and Rabiee, Hamid R.",
    booktitle = "Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers)",
    month = apr,
    year = "2025",
    address = "Albuquerque, New Mexico",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.naacl-long.464/",
    pages = "9177--9206",
}

• ManaTTS Dataset Huggingface | Github
• Link to ManaTTS Paper