eduardem/romanian-speech-v2

Research Use Only — This dataset is released strictly for personal research and educational purposes. The processing pipeline and all scripts are fully open source, but the underlying audio originates from sources with varying copyrights. Only short fragments were used under fair use provisions and EU Copyright Directive Art. 3 (text and data mining for scientific research). This dataset must not be used for redistribution of the source material, commercial purposes, or training commercially deployed models. No complete copyrighted works are reproduced — the resulting segments are short utterances (1–15 seconds) from which a model learns Romanian phonetics and prosody, not original content.

Romanian Speech v2

This dataset supersedes [romanian-speech-v1](https://huggingface.co/datasets/eduardem/romanian-speech-v1), which is now deprecated. The v1 dataset was used to fine-tune f5-tts-romanian and xtts-v2-romanian. After training, we discovered audible artifacts in both models — truncated words, misaligned sentence boundaries, and inconsistent speaker labeling — traced back to quality issues in the v1 dataset. This v2 release is a complete rebuild with a stricter pipeline: dual-pass Whisper transcription with Levenshtein validation (>= 0.96), ensemble speaker clustering (ECAPA2 + WeSpeaker), F0-based gender verification, and multi-stage audio quality filters. These changes eliminate the artifact classes found in v1.

A curated multi-speaker Romanian speech dataset for TTS model fine-tuning and speech research.

Speakers

Methodology

This section describes the full processing pipeline so that it can be replicated. All scripts are open source.

Step 0 — Cleanup and Deduplication

• —Delete junk files and directories (@eaDir, .nfo, .jpg, .txt, .DS_Store)
• —Compute SHA-256 hash for every audio file
• —Remove exact duplicates (keep first alphabetically)

Step 1 — Audio Inventory

• —Run ffprobe on every audio file to extract duration, sample rate, channels, and bitrate
• —Reject corrupt files and files shorter than 10 seconds (RSS corpus exempted — its files are pre-segmented utterances of 2–5 seconds)
• —For M4B audiobooks: extract individual chapters via ffprobe chapter metadata and inventory each separately

Step 2 — Speaker Identification

Speaker clustering uses an ensemble of two embedding models to maximize accuracy:

1. 1.ECAPA2 (Jenthe/ECAPA2) — TorchScript model, 192-dimensional embeddings, 0.34% EER on VoxCeleb
2. 2.WeSpeaker ResNet293-LM — ONNX model via onnxruntime-gpu, 256-dimensional embeddings, 0.53% EER on VoxCeleb

For each audio file, a 30–60 second sample from the middle is extracted and embedded by both models. Cosine distance matrices from each model are min-max normalized and averaged. Agglomerative clustering (average linkage) is applied with a distance threshold of 0.25 (validated on a stable plateau at 0.25–0.35).

Known speakers from public datasets (RSS, LibriVox) are assigned directly. Speaker gender is verified using F0 (fundamental frequency) analysis. Low-quality clusters (score < 60) and singleton speakers (only 1 source file) are excluded — 11 speakers dropped, 21 retained.

Step 3 — Transcription

• —Convert source audio to 16 kHz mono WAV (temporary, on local SSD)
• —Transcribe with Whisper large-v3 via faster-whisper on CUDA (float16)
• —Settings: language="ro", beam_size=5, vad_filter=True, word_timestamps=True
• —Output: per-file JSON with word-level timestamps
• —For RSS corpus: existing text transcripts from the corpus are also loaded for later comparison

Step 4 — Sentence Splitting

Segments are split at sentence boundaries using word-level timestamps from Whisper:

• —Primary method: Split at sentence-ending punctuation (., ?, !), respecting Romanian abbreviations (d-le, nr., str., prof., etc.) to avoid false splits
• —Fallback: When sentence splitting produces fewer than 30% usable segments (common with poetry or text without punctuation), fall back to Whisper's natural VAD-based segments
• —RSS corpus: Pre-segmented utterances are copied directly (already clean single sentences)
• —Filters: Reject segments shorter than 1 second or longer than 15 seconds, or with fewer than 2 words
• —Silence padding: 300 ms of silence appended to each segment

Step 5 — Transcript Validation

Every segment is re-transcribed independently with Whisper large-v3, and the new transcription is compared against the original (from Step 3/4):

1. 1.Both texts are normalized: lowercased, punctuation stripped, Romanian diacritics normalized (cedilla forms ţ/ş converted to comma-below forms ț/ș)
2. 2.Normalized Levenshtein similarity ratio is computed
3. 3.Segments with ratio < 0.96 are rejected — this catches truncated words, hallucinated text, misaligned boundaries, and segments where Whisper was uncertain
4. 4.This dual-pass approach (transcribe the full file, split, then re-transcribe segments) provides a strong consistency check that neither pass alone can achieve

Step 6 — Quality Filters

Each segment passes three audio-quality checks:

Segments failing any check are dropped. Spectral flatness is logged but not used as a hard filter.

Pipeline Funnel

7,995 source files (628 hours raw audio)
  │
  ├─ Step 1: Inventory & quality gate
  ▼
7,455 files transcribed (Step 3)
  │
  ├─ Step 4: Sentence splitting + segment extraction
  ▼
325,722 segments
  │
  ├─ Step 5: Whisper re-transcription + Levenshtein >= 0.96
  ▼
253,346 segments (after validation + speaker exclusion)
  │
  ├─ Step 6: SNR, clipping, silence filters
  ▼
247,244 segments (final dataset — 360.16 hours)

Overall acceptance rate: 75.9% of extracted segments pass all quality checks.

Dataset Structure

Sources

Legal Notice

This dataset was created for personal scientific research under the following legal basis:

• —RSS corpus: Released under CC BY-SA 4.0. Attribution: Adriana Stan, University of Edinburgh.
• —LibriVox recordings: Public domain. No restrictions.
• —Audiobook fragments: Used under the EU Copyright Directive (2019/790) Article 3, which explicitly permits text and data mining for scientific research purposes. Only short fragments (1–15 seconds) are included — no complete works are reproduced. The dataset is used to learn Romanian phonetics and prosody, not to reproduce or distribute copyrighted content.

You must not use this dataset to:

• —Redistribute or reconstruct the original copyrighted audio material
• —Train commercially deployed models
• —Distribute the dataset or derivatives beyond personal research

If you are a rights holder and have concerns, please open an issue on the repository.

Usage

from datasets import load_dataset

# Load full dataset
ds = load_dataset("eduardem/romanian-speech-v2")

# Filter by speaker
elena = ds["train"].filter(lambda x: x["speaker"] == "Elena")

# Filter by gender
female = ds["train"].filter(lambda x: x["gender"] == "female")

# Stream without full download
ds = load_dataset("eduardem/romanian-speech-v2", streaming=True)
for sample in ds["train"]:
    print(sample["speaker"], sample["text"][:80])
    break

Environment and Replication

All processing scripts are available in the project repository.

Limitations

• —Variable recording quality: Studio-recorded RSS segments are significantly cleaner than audiobook recordings, which may have room ambience or compression artifacts
• —Speaker imbalance: Mihai has 59.66 hours while Diana has only 0.01 hours (4 segments). Consider filtering to speakers with sufficient data for your use case
• —Sample rate variation: Ranges from 22.05 kHz to 48 kHz across speakers. Resample to a uniform rate before training
• —Transcription accuracy: While the Levenshtein >= 0.96 threshold catches most errors, some segments may contain minor diacritics or word-boundary inaccuracies

Acknowledgments

• —RSS corpus: Adriana Stan, SPED Lab, University of Edinburgh (paper)
• —LibriVox: Volunteer readers Cornel Nemes and Gabriela Oprea
• —Whisper: OpenAI (paper)
• —ECAPA2: Jenthe Thienpondt and Kris Demuynck (paper)

Citation

@dataset{romanian_speech_v2_2026,
  title={Romanian Speech v2: A Multi-Speaker Romanian TTS Dataset},
  author={eduardem},
  year={2026},
  publisher={HuggingFace},
  url={https://huggingface.co/datasets/eduardem/romanian-speech-v2}
}