trained_50_percents/tts.py

import os
import sys
current_dir = os.path.dirname(os.path.abspath(__file__))
print('add current dir to sys.path', current_dir)
sys.path.append(current_dir)
from sparktts.models.audio_tokenizer import BiCodecTokenizer
from transformers import AutoTokenizer, AutoModelForCausalLM
import soundfile as sf
import numpy as np
import torch
from utilities import generate_embeddings
def generate_speech(model, tokenizer, text, bicodec, prompt_text=None, prompt_audio=None, 
                   max_new_tokens=3000, do_sample=True, top_k=50, top_p=0.95, 
                   temperature=1.0, device="cuda:0"):
    """
    生成语音的函数
    
    Args:
        model: 语言模型
        tokenizer: 文本分词器
        text: 要生成语音的文本
        bicodec: BiCodecTokenizer 实例
        prompt_text: 提示文本（可选）
        prompt_audio: 提示音频数组（可选）
        max_new_tokens: 最大生成token数
        do_sample: 是否使用采样
        top_k: top-k采样参数
        top_p: top-p采样参数
        temperature: 温度参数
        device: 设备
    
    Returns:
        wav: 生成的音频波形
    """
    # 设置eos_token_id - 根据训练代码，eos_token_id = model.config.vocab_size - 1
    eos_token_id = model.config.vocab_size - 1
    print(f"EOS token ID: {eos_token_id}")
    
    # 生成输入嵌入
    embeddings = generate_embeddings(
        model=model,
        tokenizer=tokenizer,
        text=text,
        bicodec=bicodec,
        prompt_text=prompt_text,
        prompt_audio=prompt_audio
    )
    
    print("开始生成语音...")
    print(f"输入嵌入形状: {embeddings['input_embs'].shape}")
    global_tokens = embeddings['global_tokens'].unsqueeze(0)
    # 设置模型为评估模式
    print(f'embeddings dtype: {embeddings["input_embs"].dtype}')
    model.eval()
    
    with torch.no_grad():
        # 使用模型的generate方法
        generated_outputs = model.generate(
            inputs_embeds=embeddings['input_embs'],
            attention_mask=torch.ones((1, embeddings['input_embs'].shape[1]),dtype=torch.long,device=device),
            max_new_tokens=max_new_tokens,
            do_sample=do_sample,
            top_k=top_k,
            top_p=top_p,
            temperature=temperature,
            eos_token_id=eos_token_id,
            pad_token_id=tokenizer.pad_token_id if hasattr(tokenizer, 'pad_token_id') else tokenizer.eos_token_id,
            use_cache=True
        )
    print(f"generated_outputs: {generated_outputs}")
    
    print(f"生成的token数量: {generated_outputs.shape}")
    print(f"生成的token IDs: {generated_outputs.tolist()}")
    
    # 直接使用生成的token ID作为semantic tokens
    # 注意：这里生成的token ID是模型词表中的ID，不是原始tokenizer的词表
    semantic_tokens_tensor = generated_outputs[:,:-1]
    
    print(f"Semantic tokens shape: {semantic_tokens_tensor.shape}")
    
    
    print(f"Global tokens shape: {global_tokens.shape}")
    
    # 使用BiCodec解码生成音频
    with torch.no_grad():
        wav = bicodec.detokenize(global_tokens, semantic_tokens_tensor)
    
    print(f"生成的音频形状: {wav.shape}")
    return wav

device = 'cuda:0'

audio_tokenizer = BiCodecTokenizer(model_dir=current_dir, device=device)

print(audio_tokenizer)

tokenizer = AutoTokenizer.from_pretrained(current_dir, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(current_dir, trust_remote_code=True)
print(tokenizer)
print(model)

model = model.bfloat16().to(device)
model.eval()

prompt_text = "我们并不是通过物理移动手段找到星河的。"
prompt_audio_file = os.path.join(current_dir, 'kafka.wav')
prompt_audio, sampling_rate = sf.read(prompt_audio_file)

print(f"Loaded prompt audio from {prompt_audio_file}")
print(f"Original sampling rate: {sampling_rate}Hz")
print(f"Audio shape: {prompt_audio.shape}")

target_sample_rate = audio_tokenizer.config['sample_rate']
if sampling_rate != target_sample_rate:
    print(f"Resampling from {sampling_rate}Hz to {target_sample_rate}Hz...")
    from librosa import resample
    prompt_audio = resample(prompt_audio, orig_sr=sampling_rate, target_sr=target_sample_rate)
    prompt_audio = np.array(prompt_audio, dtype=np.float32)
    print(f"Resampled audio shape: {prompt_audio.shape}")
else:
    print(f"Audio sampling rate already matches target ({target_sample_rate}Hz)")

text = "为了点燃青少年对科技的热情，培养他们的创新思维与动手能力，杏花岭区巨轮街道社区教育学校携手中车社区教育分校，与太原市科学技术协会联手，于暑期精心策划了一场别开生面的青少年数智技术服务港探索之旅，吸引了众多社区青少年的积极参与。"
wav = generate_speech(model, tokenizer, text, audio_tokenizer, prompt_audio=prompt_audio, device=device)
sf.write('output.wav', wav, target_sample_rate)