Speech Synthesis (Roadmap / Experimental)

Experimental — No Trained Model Available Yet

The olaverse.speech module provides a TTS pipeline architecture, but olaverse does not yet ship a trained acoustic model or vocoder.

• Production-ready: text normalisation (TTSNormalizer) and diacritisation (Diacritizer) — these are fully functional and available in olaverse.nlp.
• Experimental: acoustic synthesis (text → Mel-spectrogram → audio waveform). The architecture is here; trained weights are on the roadmap.

Using speech classes will emit an ExperimentalWarning. To silence it:

import warnings
from olaverse import ExperimentalWarning
warnings.filterwarnings("ignore", category=ExperimentalWarning)

The diacritizers are the most valuable unfinished asset here — restoring tones from plain Yoruba text is the hardest front-end step of any Yoruba TTS system, and that part is done.

---

What works today

Use TTSNormalizer and Diacritizer directly for the NLP front-end of a TTS pipeline:

from olaverse.nlp import TTSNormalizer, Diacritizer

normalizer = TTSNormalizer(lang="yo")
diacritizer = Diacritizer(model="diacnet-yor-viterbi")

text = "Dr. Ade lo si oja lana"
normalized = normalizer.normalize(text)      # "Dọ́kítà Ade lo si oja lana"
diacritized = diacritizer.restore(normalized) # "Dọ́kítà Adé ló sí ọjà lànà"

---

TTS Pipeline Architecture

TTSPipeline wires together all four steps and is ready to use once you supply an acoustic model and vocoder. Steps 1–2 work now; steps 3–4 require your own models.

from olaverse import TTSPipeline

# Steps 1 & 2 (normalisation + diacritisation) work without custom models
pipeline = TTSPipeline(lang="yo")
result = pipeline.synthesize("Mr. Ade lo si oja lana")

print(result["normalized_text"])   # "Míṣìtà Ade lo si oja lana"
print(result["diacritized_text"])  # "Míṣìtà Adé ló sí ọjà lànà"
print(result["audio"])             # None — no acoustic model provided
print(result["status"])            # "Acoustic model or Vocoder not provided."

Injecting your own models

If you have a trained acoustic model and vocoder, implement the base classes and inject them:

from olaverse import TTSPipeline, BaseAcousticModel, BaseVocoder

class MyAcousticModel(BaseAcousticModel):
    def load_weights(self, path):
        ...
    def forward(self, text):
        ...  # returns Mel-spectrogram

class MyVocoder(BaseVocoder):
    def load_weights(self, path):
        ...
    def generate(self, mel):
        ...  # returns audio waveform

pipeline = TTSPipeline(
    lang="yo",
    acoustic_model=MyAcousticModel(),
    vocoder=MyVocoder(),
)
result = pipeline.synthesize("Ẹ káàárọ̀")
# result["audio"] now contains the waveform

---

Model Interfaces

olaverse.speech.TTSPipeline

TTSPipeline(lang: str = 'yo', acoustic_model: BaseAcousticModel = None, vocoder: BaseVocoder = None, diacritizer_model: str = 'diacnet-yor-viterbi')

End-to-end Text-to-Speech pipeline architecture.

Orchestrates: Text Normalisation → Tone Restoration → Acoustic Model → Vocoder.

.. warning:: Experimental — no trained acoustic model or vocoder available yet. Steps 1 (normalisation) and 2 (diacritisation) are fully functional. Steps 3 and 4 require you to inject your own acoustic model and vocoder. End-to-end audio synthesis from olaverse is on the roadmap.

To silence this warning:
    import warnings
    warnings.filterwarnings("ignore", category=ExperimentalWarning)

Functions

synthesize

synthesize(text: str)

Synthesise raw text into an audio waveform.

Returns a dict with keys

• normalized_text: text after abbreviation/number expansion
• diacritized_text: text after tone restoration
• audio: waveform array/tensor, or None if no acoustic model/vocoder provided
• status: "Success" or a message explaining what is missing

olaverse.speech.BaseAcousticModel

Bases: ABC

Abstract base class for acoustic models (e.g. FastSpeech, Tacotron). Converts normalised/diacritised phonetic text into Mel-spectrograms.

.. warning:: Experimental — no trained model available yet. Subclassing this is fine for custom integrations, but olaverse does not yet ship a trained acoustic model. This is on the roadmap.

Functions

load_weights abstractmethod

load_weights(path: str)

Load PyTorch/ONNX model weights from the specified path.

forward abstractmethod

forward(text: str)

Convert text into acoustic features (e.g. a Mel-spectrogram tensor).

Parameters:

Name	Type	Description	Default
text	str	Phonetically normalised and diacritised text.	required

Returns: Acoustic features tensor.

olaverse.speech.BaseVocoder

Bases: ABC

Abstract base class for vocoders (e.g. HiFi-GAN, WaveGlow). Converts Mel-spectrograms into raw audio waveforms.

.. warning:: Experimental — no trained model available yet. Subclassing this is fine for custom integrations, but olaverse does not yet ship a trained vocoder. This is on the roadmap.

Functions

load_weights abstractmethod

load_weights(path: str)

Load PyTorch/ONNX model weights from the specified path.

generate abstractmethod

generate(acoustic_features: object) -> object

Convert acoustic features into a raw audio waveform.

Parameters:

Name	Type	Description	Default
acoustic_features	object	Output from a BaseAcousticModel.	required

Returns: Audio waveform array/tensor.