YoruTTS-0.5: Pretrained Yorùbá TTS Model

• The paper introduces YoruTTS-0.5, a pretrained text-to-speech model for Standard Yorùbá that demonstrates the viability of neural TTS for low-resource tonal languages.
• It employs a VITS-based architecture within Coqui TTS and uses the AcoustAug technique to expand a 1,504-sample base to 12,032 augmented recordings.
• Objective evaluations show improved pitch alignment (F0_RMSE reduced from 72.85 Hz to 63.54 Hz) and intelligible, natural outputs despite grapheme-to-phoneme modeling challenges.

YoruTTS-0.5 is a pretrained text-to-speech (TTS) model for Standard Yorùbá, built as a proof-of-concept using the BENYO-S2ST-Corpus-1, the largest public bilingual English–Yorùbá speech dataset to date. Released with open-source training code and pipelines in Coqui TTS, YoruTTS-0.5 demonstrates the viability of modern neural TTS (specifically VITS-based systems) for low-resource, tonal African languages. The model achieves moderate pitch alignment with human reference, and accessible outputs rated as intelligible and natural by human assessors, despite persisting challenges in explicit grapheme-to-phoneme modeling and voice diversity (Adetiba et al., 12 Jul 2025).

1. Data Resources and Preprocessing

BENYO-S2ST-Corpus-1 Construction

YoruTTS-0.5 is grounded in the BENYO-S2ST-Corpus-1, which provides 24,064 paired audio samples: 12,032 each in English and Yorùbá, with a total duration of 41.20 hours (17.81 h English, 23.39 h Yorùbá). English audios are synthesized using Meta AI’s MMS TTS model (“facebook/mms-tts”) from 1,504 Standard English transcripts (16 kHz, up-sampled to 22.05 kHz), while Yorùbá audios are real human recordings from the YORULECT corpus (48 kHz, down-sampled to 22.05 kHz). The original, small-scale YORULECT corpus comprises 1,504 parallel samples without paired English audio; parallelization was achieved via synthesis.

Acoustic Data Augmentation: AcoustAug

To address low-resource constraints and enhance generalization, the AcoustAug algorithm augments each original 22.05 kHz audio by applying latent acoustic feature transformations: speed scaling (factors 0.9, 1.1), pitch shifting (0.95, 1.05), and loudness variation (−5 dB, +5 dB, +10 dB). Each original yields eight distinct versions, expanding 1,504 raw audios to 12,032 per language. Objective pre-augmentation checks using STOI, PESQ, and LogSpec-L1 confirmed the preservation of spectral fidelity within moderate augmentation bounds.

2. Model Architecture

Core Structure

YoruTTS-0.5 is an adaptation of the Variational Inference Text-to-Speech (VITS) architecture, implemented in the open-source Coqui TTS framework. The architecture integrates:

• A posterior encoder $q(z|x)$ mapping real speech $x$ to latent $z$
• A prior encoder $p(z|c)$ mapping tokenized Yorùbá text $c$ to $z$
• Normalizing flows $f$ transforming a simple Gaussian prior to a more complex latent distribution
• A HiFi-GAN-style decoder $p_w(y|z)$ that synthesizes waveforms $y$ from $z$

Text Processing and Acoustic Configuration

• Text processing omits explicit grapheme-to-phoneme mapping (“use_phonemes=False”), relying on custom CharactersConfig for Yorùbá graphemes and diacritics, UTF-8 normalization, and tokenization via TTSTokenizer.
• Acoustic parameters: sample rate 22,050 Hz, n_fft=1,024, hop_length=256, 80 mel filters, mel_fmin=0 Hz, mel_fmax=None.

3. Training Regimen

YoruTTS-0.5 was trained for 1,000 epochs using the Adam optimizer (β₁=0.8, β₂=0.99) at a learning rate of $x$0, consistent with default Coqui VITSConfig settings (typical batch_size=32), and employed mixed-precision training (float16/float32). The loss structure is ELBO-styled: mel-spectrogram reconstruction loss ($x$1), KL divergence ($x$2), adversarial generator/discriminator losses ($x$3, $x$4), feature-matching loss ($x$5), and duration prediction loss ($x$6).

4. Objective Evaluation and Results

Fundamental Frequency RMSE

The principal metric for evaluating pitch accuracy is the fundamental-frequency root mean squared error (F0_RMSE), defined as: $x$7 where $x$8 is the number of frames, $x$9 is the predicted pitch, and $z$0 is the reference pitch.

YoruTTS-0.5 achieves:

• F0_RMSE = 72.85 Hz after 540 epochs
• F0_RMSE = 63.54 Hz after 1,000 epochs

These results indicate moderate alignment of synthesized pitch contours with reference human audio.

Perceptual Quality

Human listener evaluations reported that the synthesized outputs are natural and intelligible, despite the moderate absolute pitch deviations. This suggests perceptual quality in TTS is affected by multiple correlates—not solely F0 error—but also by prosody, timbre, and rhythm.

5. Strengths and Limitations

Advantages

• First pretrained Yorùbá TTS released using Coqui, validated on a >40 h bilingual corpus.
• The AcoustAug augmentation pipeline enhances robustness in low-resource regimes.
• Modular, open-source pipeline with resources and code accessible on HuggingFace.

Constraints

• Only single-speaker data per language, yielding limited speaker and style diversity.
• Absence of explicit Yorùbá-specific G2P modeling; the model must learn the grapheme-to-phoneme mapping implicitly.
• F0 modeling exhibits residual inaccuracies (F0_RMSE ≈ 63 Hz).

A plausible implication is that expansion to multi-speaker and multi-dialect datasets, combined with explicit phoneme conditioning and improved prosody modules, could ameliorate current limitations.

6. Future Directions and Broader Significance

Advancements proposed include integration of high-quality Yorùbá grapheme-to-phoneme tools, addition of multi-speaker/multi-dialect recordings, and exploration of explicit pitch and prosody encoders to reduce pitch RMSE. Automating corpus creation through “agentic AI” is also suggested as a pathway for cost-effective, large-scale expansion.

YoruTTS-0.5 demonstrates the feasibility of modern neural TTS methods for tonal, low-resource African languages. By providing models and data, it facilitates downstream work in Yorùbá automatic speech recognition, speech-to-text, and direct speech-to-speech translation, and offers a reproducible template for similar efforts in other low-resource high-value language pairs (Adetiba et al., 12 Jul 2025).