FraPiz/moldovan-dialectal-romanian-speech-corpus

Moldovan Dialectal Romanian Educational Speech Corpus

This dataset contains aligned Romanian educational speech with Moldovan dialectal characteristics. It was constructed from publicly accessible lesson videos recorded by teachers from the Republic of Moldova and published through the EducatieOnline platform.

The corpus supports research on automatic speech recognition (ASR), text-to-speech synthesis (TTS), forced alignment, and low-resource dialectal speech processing.

Dataset summary

Statistic	Value
Validated audio-text segments	35,760
Total duration	70.673842 h
Mean segment duration	7.1148 s
Median segment duration	6.64 s
Minimum duration	2.00 s
Maximum duration	15.00 s
Inferred speakers/teachers	56
Source lesson videos	418
Disciplines	11
Grade levels	3
Word tokens	491,019
Lowercased token types	35,296

Data structure

The repository contains:

metadata.jsonl
data/
`-- train-00000-of-00036.parquet

The Parquet shards embed the WAV bytes and expose the following fields:

{
  "audio": {"bytes": "<WAV bytes>", "path": "example.wav"},
  "text": "normalized transcript",
  "duration": 7.12,
  "audio_filename": "example.wav"
}

The original workflow manifest is named train.jsonl. For Hub compatibility, this repository provides an equivalent auxiliary metadata.jsonl with stable audio filenames, while the actual audio is embedded in 36 Parquet shards. The manifest contains the complete released corpus, not only the experimental training partition. Hugging Face therefore exposes it as one distribution split named train.

Experimental splits

The associated ASR experiments shuffled the complete corpus at segment level with seed 42 and used an 80/10/10 train/validation/test protocol. The held-out ASR test partition contained 3,576 segments and approximately 6.99 hours.

The TTS preparation independently excluded segments outside the configured 1.00-11.61-second range and retained 31,472 utterances (55.05 hours). With seed 42, the resulting split contained 25,178 training utterances (44.06 hours), 3,147 evaluation utterances (5.47 hours), and 3,147 test utterances (5.52 hours).

The published manifest does not include an explicit split column. Exact experimental reproduction therefore requires applying the procedure and seed described above or using the split manifests released with the corresponding training code.

Construction workflow

1. Educational lesson videos were collected from EducatieOnline.
2. Audio was extracted from the source videos.
3. Recordings containing background music were filtered or cleaned.
4. File-level transcripts were generated before segmentation.
5. Romanian Wav2Vec2 CTC forced alignment and Silero VAD were used to identify speech intervals.
6. Segment boundaries were refined using an RMS-energy profile and constrained to 2-15 seconds.
7. Audio-text consistency and manifest/file correspondence were validated.
8. Metadata encoded in source identifiers was retained in the audio filenames.

The final release contains exactly one WAV file for every JSONL record.

Source and provenance

• EducatieOnline platform: https://www.educatieonline.md/
• Source lesson collection: https://drive.google.com/drive/folders/1OTsY6YhnzKGuFnFmJqLPvhO5onr1E3zD
• Original corpus release: https://drive.google.com/drive/folders/1Va0tmpp-Q8A2gGbFwLignYVtf0KD6Uth

Audio filenames encode source metadata such as grade, discipline, teacher name, lesson title, source identifier, and segment index. Consequently, the dataset contains identifiable teacher names and voice recordings. It is not an anonymized speech corpus.

Intended uses

• Romanian and Moldovan-dialect ASR research.
• Dialectal and low-resource TTS adaptation.
• Speech-text alignment and segmentation research.
• Analysis of educational speech and domain adaptation.

Limitations

• The corpus is dominated by planned educational speech and does not represent spontaneous conversation.
• Speaker contributions are uneven.
• The experimental segment-level split is not speaker-independent or source-video-independent.
• Automatically inferred speaker-gender metadata is approximate and is not included as a ground-truth annotation.
• Transcripts and alignments may still contain residual automatic-processing errors despite validation.
• The dataset should not be treated as representative of every Moldovan or Romanian dialectal variety.

Rights, privacy, and responsible use

No standard open-content license is asserted for the source lesson recordings. The source content and speaker rights remain subject to the rights and terms of the original platform, educators, and other applicable rightsholders. Users are responsible for establishing an appropriate legal basis for access, redistribution, biometric processing, model training, publication, and any commercial use.

Do not use the dataset or derived models for unauthorized speaker identification, impersonation, surveillance, deception, or harmful biometric profiling. Public availability of the source lessons does not by itself remove privacy, personality, copyright, or data-protection obligations.

Authors and project

Marius Cerescu, Alexandr Parahonco, Olesea Caftanatov, Tudor Bumbu, Nichita Degteariov, and Ion Bostan. Vladimir Andrunachievici Institute of Mathematics and Computer Science, Moldova State University, Chisinau, Republic of Moldova.

This work was elaborated within the project Platforma Educationala bazata pe Inteligenta Artificiala "Guguta", registered in the State Register of projects in science and innovation, code 25.80012.0807.21TC, project leader Tudor Bumbu, Dr.

Citation

@inproceedings{cerescu2026moldovan,
  title     = {Fine-Tuning Neural Speech Models for Romanian Speech Recognition and Synthesis with Moldovan Dialectal Data},
  author    = {Cerescu, Marius and Parahonco, Alexandr and Caftanatov, Olesea and Bumbu, Tudor and Degteariov, Nichita and Bostan, Ion},
  booktitle = {International Conference on System Analysis and Intelligent Information Technologies (SAIIT)},
  year      = {2026}
}