OmniVoice

Generates natural speech from text input across 12+ languages without requiring language-specific fine-tuning or training data. The model uses a unified encoder-decoder architecture that learns language-agnostic phonetic and prosodic representations, enabling it to synthesize speech in any supported language by conditioning on language tokens and text embeddings. This approach eliminates the need for separate language-specific models or extensive multilingual training datasets.

Enables synthesis of speech in a target speaker's voice by extracting speaker embeddings from a short reference audio sample (typically 5-30 seconds) and conditioning the decoder on these embeddings. The model uses speaker-agnostic phonetic encodings combined with speaker-specific prosodic and timbre information, allowing zero-shot voice cloning without speaker-specific training. This is implemented via speaker embedding extraction (using a pre-trained speaker encoder) and adaptive layer normalization in the decoder.

Converts input text into phoneme sequences and extracts linguistic features (stress, tone, syllable boundaries) that condition the speech synthesis decoder. The model uses a language-specific grapheme-to-phoneme (G2P) converter or pre-computed phoneme mappings, combined with linguistic feature extractors that identify prosodic boundaries and emphasis patterns. This enables the model to generate speech with accurate pronunciation and natural prosody without explicit prosody annotations.

Supports both batch synthesis (processing multiple text inputs simultaneously) and streaming synthesis (generating audio incrementally as text becomes available). The implementation uses a sliding window decoder that processes phoneme sequences in chunks, enabling low-latency streaming while maintaining prosodic coherence across chunk boundaries. Batch processing leverages GPU parallelization to synthesize multiple utterances concurrently, with adaptive buffering to manage memory constraints.

Uses the safetensors format for model storage, enabling fast and secure model loading with built-in integrity verification. Safetensors is a binary format that stores model weights with explicit type information and checksums, allowing the model to be loaded directly into GPU memory without intermediate Python object deserialization. This approach reduces model loading time by 30-50% compared to PyTorch pickle format and eliminates arbitrary code execution risks during model deserialization.

Uses a universal phonetic encoder that maps phoneme sequences from any supported language into a shared acoustic feature space, combined with language-specific decoder branches that generate speech acoustics tailored to each language's phonological and prosodic characteristics. The encoder learns language-agnostic representations through contrastive learning across multilingual phoneme pairs, while decoder branches capture language-specific spectral and temporal patterns. This hybrid approach enables zero-shot synthesis while maintaining language-specific acoustic quality.

Converts mel-spectrogram outputs from the acoustic model into high-quality audio waveforms using a pre-trained neural vocoder (typically HiFi-GAN or similar architecture). The vocoder uses dilated convolutions and residual connections to upsample spectrograms to waveform resolution while maintaining spectral fidelity. The integration is modular, allowing different vocoders to be swapped without retraining the acoustic model, enabling trade-offs between audio quality and inference latency.