zuhri025/lm_weights_small_tts_prepared

• <|audio_0|> … <|audio_12801|> — audio codebook (12,802 tokens, IDs 50011–62812)
• Boundary: <|start_of_audio|>, <|end_of_audio|>, <|audio|>
• Text gates: <|start_of_text|>, <|end_of_text|>
• Conditioning: <|male|>, <|female|>, <|other|>, <|accent_us|>, <|accent_uk|>, <|unconditional|>

Embedding init

• Special/structural tokens → mean of original embeddings + σ=0.02 noise
• Audio code tokens → σ=0.02 random normal (independent codebook)
• All original embeddings unchanged.