NAACL · 2025 · Conference
Fetrat Qharabagh et al. (Sharif University of Technology) · → Paper · Demo: ? · Code: ✓
ManaTTS is a publicly released, CC-0-licensed Persian TTS corpus of 86 hours from a single speaker, accompanied by a fully open pipeline that handles forced alignment under text-audio mismatch conditions and provides a template for building TTS datasets in other low-resource languages.
Problem
Persian TTS research is constrained by the lack of open, large-scale, single-speaker corpora. Existing Persian datasets are either not publicly available, carry restrictive or unspecified licenses, are limited in duration (most under 30 hours), are domain-restricted (several are drawn exclusively from Quran recitations), or contain synthesized rather than natural speech. A dataset that satisfies all criteria simultaneously — open license, natural speech, large size, topical diversity, and a reproducible creation pipeline — did not exist before this work.
Method
The dataset is sourced from the Nasl-e-Mana monthly magazine, a Persian publication for the blind community, whose owners granted a CC-0 license. Audio files are first preprocessed by converting from MP3 to WAV and running source separation (Spleeter) to remove background music. Text files are normalized using the Hazm library, with numbers converted to spoken equivalents and links stripped.
The central contribution of the pipeline is a two-stage alignment system that tolerates approximate text-audio correspondence. In the first stage, start-end alignment trims audio to match its corresponding text by sliding a window over silence-detected segments and selecting the trim that minimises CER against a reference. The second stage, forced alignment, segments the trimmed pairs into 2-12 second chunks. Because standard tools like Aeneas require exact text-audio correspondence, the authors developed a custom forced alignment algorithm that uses two search strategies: Interval Search (contiguous text substring matching) and Gapped Search (allowing a gap in the reference text to handle censored or corrected utterances). A chunk is accepted if CER falls below 0.2, with an immediate acceptance threshold of 0.05.
Critically, hypothesis transcripts for alignment come from a majority-of-experts transcription module that aggregates five open-source Persian ASR models (Vosk, Wav2vec-v3, Wav2vec-fa, Whisper-fa, Hezar) and discards transcripts shorter than 80% of the longest, filtering truncation errors. The ranking of ASR models is calibrated on PersianInformal, a small evaluation dataset collected for this work. The pipeline produced 64,834 accepted audio-text chunks from 564 source files, with a 97.98% acceptance rate. The final corpus has 44.1 kHz sampling rate, 24,113 unique words, and spans three years of magazine issues.
To validate corpus quality, a Tacotron2-based multi-speaker model is trained on ManaTTS, using a speaker encoder, a sequence-to-sequence Tacotron2 mel predictor, and a pre-trained HiFi-GAN vocoder. Resampling to 24 kHz is applied for training, with 80 mel filter banks and an FFT size of 2048. Training runs for 320,000 steps across six sessions with learning rate decay from 1e-3 to 1e-5. The model is not the primary contribution; it serves as a downstream quality probe.
Key Results
A subjective MOS test with 76 native Persian speakers evaluated five held-out utterances across five conditions. The Tacotron2 model trained on ManaTTS achieved MOS 3.76, compared to MOS 3.86 for the HiFi-GAN vocoder applied to ground-truth mel-spectrograms, and MOS 4.01 for natural speech. The two open-access baseline Persian TTS models (VITS-based and Glow-TTS-based) scored 1.68 and 1.34 respectively, substantially below the proposed system (§5.1, Table 2). On objective metrics (PESQ, MCD, CER via ASR), the proposed system also outperforms both baselines on PESQ (1.11 vs. 1.05 and 1.06) and achieves lower CER by Google Speech API (0.096 vs. 0.126 and 0.233) (§D.2, Tables 9-10).
The dataset’s transcription accuracy was validated by manually transcribing 100 random chunks: the CER between manual and pipeline-generated transcriptions was 0.01, indicating near-perfect text-audio correspondence despite the automated pipeline (§5.2).
The robustness of the multi-ASR majority-voting alignment approach was tested by injecting artificial character errors into ASR outputs. At a 25% additional CER degradation, the multi-ASR system rejected 9 out of 151 chunks, while a single-ASR approach rejected 65, demonstrating the alignment system’s tolerance for weak individual models (§C, Table 7).
Novelty Assessment
The contribution is primarily a dataset and an associated processing pipeline. The forced alignment algorithm, which accommodates imperfect text-audio correspondence via Interval and Gapped Search strategies combined with a multi-ASR transcription module, represents a genuine engineering contribution adapted to the low-resource language setting. The approach extends naturally to other languages where moderate-quality ASRs are available but text-audio matching is noisy. The TTS model trained on the corpus is a standard Tacotron2 architecture and is not presented as a methodological contribution; its purpose is to demonstrate corpus usability.
The key limitation of the validation is that comparisons are against pre-existing open-access Persian TTS baselines that score poorly (MOS below 1.7), which makes the proposed corpus’s utility appear strong but does not speak to whether ManaTTS reaches the quality of well-studied English or Mandarin corpora.
Field Significance
Moderate — ManaTTS is the largest open-access single-speaker Persian TTS dataset at the time of publication, filling a concrete gap in Persian speech technology infrastructure. Its fully reproducible pipeline, designed to tolerate near-miss text-audio alignment, provides a reusable methodology for corpus construction in other low-resource languages with limited high-quality ASR support.
Claims
-
supports: Open-pipeline corpus construction using multi-ASR majority voting can produce high-quality aligned TTS data even when no single ASR model is reliable.
Evidence: The ManaTTS pipeline achieved 97.98% chunk acceptance and a manual CER of 0.01 on 100 sampled chunks, despite relying on Persian ASRs with 10-30% CER on evaluation sets. (§5.2, §C, Table 7)
-
supports: The absence of open-access TTS corpora is a primary bottleneck for speech synthesis in low-resource languages.
Evidence: All prior Persian TTS datasets are either not publicly available, synthetically generated, domain-limited (Quran-only), or under restrictive licenses; ManaTTS is the first to address all five criteria simultaneously. (§2, Table 1, §6)
-
complicates: High MOS scores on a TTS system trained in a low-resource setting may reflect weak baselines rather than strong absolute quality.
Evidence: The proposed Tacotron2 model scores MOS 3.76, outperforming open-access Persian VITS and Glow-TTS baselines that score 1.68 and 1.34, respectively; the gap to natural speech (4.01) and to GT-spectrogram synthesis (3.86) is non-trivial. (§5.1, Table 2)
-
refines: Forced alignment pipelines for TTS corpus construction require robustness to text-audio mismatch when sourcing from naturalistic recordings.
Evidence: Standard alignment tools (e.g., Aeneas) assume exact text-audio correspondence; the custom gapped-interval alignment algorithm accepted 99.39% of chunks via Interval Search and handled mismatch cases via Gapped Search, critical for magazine-sourced data where speakers censor or self-correct. (§3.2.3, Figure 11)
Limitations and Open Questions
Warning
The MOS evaluation uses only five utterances and a single speaker, and comparisons are restricted to open-access Persian baselines that score far below acceptable naturalness. The evaluation does not benchmark against current state-of-the-art multilingual TTS systems, so the dataset’s ceiling for trained model quality remains unknown.
The pipeline lacks explicit handling for English-embedded text (common in Persian), which leads to forced mismatches and chunk rejection for any English substrings. Number normalisation is incomplete for certain formats (phone numbers, dates) where the spoken form diverges from written conventions. Both limitations are acknowledged by the authors and represent concrete improvement areas for future pipeline versions (§8).
The dataset is single-speaker and single-domain (a magazine for the blind community), limiting stylistic and prosodic diversity for downstream training. Voice impersonation risk is noted in the datasheet but no anonymisation is provided.
Wiki Connections
- Evaluation Metrics — The paper employs MOS, CER, PESQ, and MCD to evaluate both corpus quality and trained model performance, with manual CER verification as the ground-truth quality signal.
- Subjective Evaluation — The MOS study involves 76 native Persian speakers evaluating five utterances across five speech sources, with randomised presentation order to control for sequence bias.