{
    "_name_or_path": "LuisChary/LatinX-TTS",
    "model_type": "latinx_tts",
    "architectures": [
        "LatinXTTS"
    ],
    "library_name": "latinx-inference",
    "pipeline_tag": "text-to-speech",
    "task": "text-to-speech",
    "license": "cc-by-nc-4.0",
    "paper": {
        "title": "LatinX: Aligning a Multilingual TTS Model with Direct Preference Optimization",
        "arxiv": "2509.05863"
    },
    "authors": [
        "Luis Felipe Chary",
        "Miguel Arjona Ramirez"
    ],
    "description": "LatinX is a multilingual, voice-preserving text-to-speech model component designed for cascaded speech-to-speech translation research.",
    "supported_languages": [
        "pt",
        "en",
        "es",
        "fr",
        "it",
        "ro"
    ],
    "language_names": {
        "pt": "Portuguese",
        "en": "English",
        "es": "Spanish",
        "fr": "French",
        "it": "Italian",
        "ro": "Romanian"
    },
    "audio": {
        "sampling_rate": 16000,
        "channels": 1,
        "format": "waveform"
    },
    "inference": {
        "repository": "https://github.com/luischary/latinx-inference",
        "entrypoint": "scripts/run_inference.py",
        "recommended_usage": "Use the accompanying latinx-inference repository rather than loading this repository directly with a generic Transformers pipeline.",
        "supports_reference_audio_conditioning": true,
        "recommended_mode": "reference_audio_conditioning"
    },
    "components": {
        "text_normalization": {
            "description": "Language-aware text normalization, including numeric value expansion in the inference repository."
        },
        "g2p": {
            "description": "Grapheme-to-phoneme conversion module.",
            "checkpoint": "g2p.pt"
        },
        "phoneme_tokenizer": {
            "description": "Phoneme-level tokenization used before acoustic token generation."
        },
        "acoustic_model": {
            "description": "Autoregressive decoder-only acoustic-token generator.",
            "checkpoint": "ar_decoder.pt",
            "parameter_count": "approximately 210M",
            "context_length": 8192
        },
        "acoustic_codec": {
            "description": "VQ acoustic autoencoder used to decode generated acoustic tokens.",
            "checkpoint": "vq_autoencoder.pt",
            "mel_bins": 80,
            "temporal_downsampling": 4,
            "codebook_size": 4096,
            "nominal_bitrate_kbps": 7.5
        },
        "vocoder": {
            "description": "HiFi-GAN waveform synthesizer.",
            "checkpoint": "hifi.pt"
        }
    },
    "checkpoint_files": [
        "g2p.pt",
        "ar_decoder.pt",
        "vq_autoencoder.pt",
        "hifi.pt"
    ],
    "training": {
        "alignment_method": "Direct Preference Optimization",
        "training_code_included": false
    },
    "limitations": {
        "long_form_generation": false,
        "requires_sentence_chunking_for_long_text": true,
        "unconditioned_generation_may_be_less_stable": true,
        "pronunciation_and_prosody_may_vary_by_language_speaker_and_domain": true
    },
    "responsible_use": {
        "requires_consent_for_reference_voice": true,
        "synthetic_speech_disclosure_recommended": true,
        "disallowed_uses": [
            "impersonation",
            "deceptive voice cloning",
            "non-consensual voice cloning",
            "fraud",
            "harassment",
            "misinformation",
            "commercial use"
        ]
    }
}