higgs-audio-v2-generation-3B-base

Generates natural-sounding speech from text input using a 3B-parameter transformer-based encoder-decoder architecture trained on multilingual corpora. The model processes tokenized text through a learned embedding space and decodes into mel-spectrogram representations, which can be converted to waveforms via vocoder integration. Supports English, Mandarin Chinese, German, and Korean with language-specific phoneme handling and prosody modeling.

Converts raw text input into phoneme sequences and linguistic features (stress, tone, duration markers) specific to each supported language before feeding to the transformer encoder. Implements language-specific text normalization (number-to-word conversion, abbreviation expansion, punctuation handling) and phoneme inventory mapping for English, Mandarin (with tone markers), German, and Korean (Hangul decomposition). This preprocessing ensures the model receives structurally consistent linguistic representations across languages.

The transformer decoder generates variable-length mel-spectrogram frames conditioned on phoneme embeddings, with auxiliary heads predicting frame duration and fundamental frequency (pitch) contours. Duration prediction enables the model to learn natural speech timing (e.g., longer vowels, shorter consonants) without explicit alignment annotations, while pitch prediction captures prosodic variation (intonation, stress patterns). The architecture uses attention mechanisms to align phonemes to acoustic frames dynamically.

The model outputs mel-spectrogram representations (80-dimensional frequency bins) that are decoupled from any specific vocoder, allowing downstream integration with multiple neural vocoder backends (HiFi-GAN, Glow-TTS vocoder, WaveGlow, etc.). This design enables users to swap vocoders based on quality/speed tradeoffs without retraining the TTS model. The mel-spectrogram format is a standard intermediate representation in speech synthesis, ensuring compatibility with existing vocoder ecosystems.

Implements a sequence-to-sequence transformer architecture where the encoder processes phoneme embeddings and the decoder generates mel-spectrogram frames using cross-attention over encoder outputs. The cross-attention mechanism learns to align phonemes to acoustic frames dynamically, enabling the model to handle variable-length inputs and outputs. The architecture uses standard transformer components (multi-head attention, feed-forward networks, layer normalization) scaled to 3B parameters with optimizations for inference efficiency.

Supports inference in four languages (English, Mandarin Chinese, German, Korean) with language-specific preprocessing and model routing. The model can accept a language code parameter to apply the correct text normalization, phoneme inventory, and linguistic feature extraction for each language. This enables building multilingual applications that either require explicit language specification or can auto-detect language from input text and route to the appropriate preprocessing pipeline.

The model is distributed via HuggingFace Hub using the safetensors format (a safer, faster alternative to pickle-based PyTorch checkpoints) with 295K+ downloads, enabling easy model loading via the transformers library. The Hub integration provides automatic model versioning, commit history, model card documentation, and community discussion features. Users can load the model with a single line of code: `AutoModel.from_pretrained('bosonai/higgs-audio-v2-generation-3B-base')`, which handles weight downloading, caching, and device placement.

The model is released as open-source under a permissive license (marked as 'other' on HuggingFace, likely Apache 2.0 or MIT based on bosonai's typical licensing), enabling free use for commercial applications, research, and fine-tuning without licensing fees or usage restrictions. The open-source release includes model weights, architecture details (via arXiv paper 2505.23009), and community access for contributions, bug reports, and improvements.