Cracking the Code: Speech Synthesis for Tonal Languages

In English, changing your pitch might change the emotion of a sentence, but it rarely changes the definition of a word. In tonal languages like Yoruba or Igbo, pitch is definition. The word 'Owo' can mean 'Money', 'Hand', or 'Honor' depending entirely on the pitch contour (High-High, Mid-Mid, Low-Low).

The Spectrogram Challenge

Standard Text-to-Speech (TTS) models like Tacotron or FastSpeech are optimized for non-tonal or weakly tonal languages. They tend to smooth out pitch curves, resulting in "robotic" speech that, for a Yoruba speaker, renders the output meaningless or comical.

When building Dára, we realized we couldn't just fine-tune existing architectures. We needed a fundamental shift in how the model predicts prosody.

Our Approach: Explicit Pitch Embeddings

Instead of letting the model implicitly learn pitch from the audio, we explicitly feed pitch information as a dedicated input feature. We modified the transformer architecture to include a "Tone Encoder."

• We developed a custom phonemizer that preserves tone diacritics.
• We trained on a proprietary dataset of high-fidelity recordings from native speakers who were instructed to exaggerate tonal distinctions.
• We utilized a Variational Autoencoder (VAE) to model the variance in pitch, allowing Dára to generate expressive, natural-sounding speech that respects semantic meaning.

The result is Dára Voice Core: a system that doesn't just read text, it understands the melody of the language. This is crucial for gaining trust. A home assistant that mispronounces your name or your language is a novelty; one that speaks like an elder is a companion.