對話音頻生成
受到一段YouTube视频的启发，该视频讨论了DeepSeek-V3，我一直在尝试使用AI生成对话。我的目标是利用Google文本转语音和ffmpeg进行音频生成和拼接，创造出逼真的音频对话。以下代码概述了我目前模拟自然来回对话的方法。
提示
创建一个自然且延长的对话，至少100轮，由两位专家A和B进行。专家们应深入讨论一个特定话题，对话应来回流动。两位参与者都应提出问题、分享见解，并探讨主题的细微差别。格式如下：
[
{
"speaker": "A",
"line": "嘿，我最近经常听到关于机器学习（ML）、深度学习（DL）和GPT的讨论。你能给我解释一下吗？"
},
{
"speaker": "B",
"line": "当然！我们从基础开始。机器学习是计算机科学的一个领域，系统通过数据学习以提高性能，而无需明确编程。可以把它想象成教计算机识别模式。"
}
]
代码
import os
import json
import random
import subprocess
from google.cloud import texttospeech
import tempfile
import time
import argparse
# 固定输出目录用于对话
OUTPUT_DIRECTORY = "assets/conversations"
INPUT_DIRECTORY = "scripts/conversation"
def text_to_speech(text, output_filename, voice_name=None):
print(f"生成音频：{output_filename}")
try:
client = texttospeech.TextToSpeechClient()
synthesis_input = texttospeech.SynthesisInput(text=text)
voice = texttospeech.VoiceSelectionParams(language_code="en-US", name=voice_name)
audio_config = texttospeech.AudioConfig(
audio_encoding=texttospeech.AudioEncoding.MP3,
effects_profile_id=["small-bluetooth-speaker-class-device"]
)
retries = 5
for attempt in range(1, retries + 1):
try:
response = client.synthesize_speech(input=synthesis_input, voice=voice, audio_config=audio_config)
with open(output_filename, 'wb') as out:
out.write(response.audio_content)
print(f"音频内容写入 {output_filename}")
return True
except Exception as e:
print(f"尝试 {attempt} 时出错：{e}")
if attempt == retries:
print(f"经过 {retries} 次尝试后，音频生成失败。")
return False
wait_time = 2 ** attempt
print(f"等待 {wait_time} 秒后重试...")
time.sleep(wait_time)
except Exception as e:
print(f"生成音频时发生错误 {output_filename}：{e}")
return False
def process_conversation(filename):
filepath = os.path.join(INPUT_DIRECTORY, filename)
output_filename = os.path.join(OUTPUT_DIRECTORY, os.path.splitext(filename)[0] + ".mp3")
if os.path.exists(output_filename):
print(f"音频文件已存在：{output_filename}")
return
try:
with open(filepath, 'r', encoding='utf-8') as f:
conversation = json.load(f)
except Exception as e:
print(f"加载对话文件 {filename} 时出错：{e}")
return
temp_files = []
voice_options = ["en-US-Journey-D", "en-US-Journey-F", "en-US-Journey-O"]
voice_name_A = random.choice(voice_options)
voice_name_B = random.choice(voice_options)
while voice_name_A == voice_name_B:
voice_name_B = random.choice(voice_options)
for idx, line_data in enumerate(conversation):
speaker = line_data.get("speaker")
line = line_data.get("line")
if not line:
continue
temp_file = os.path.join(OUTPUT_DIRECTORY, f"temp_{idx}.mp3")
temp_files.append(temp_file)
voice_name = None
if speaker == "A":
voice_name = voice_name_A
elif speaker == "B":
voice_name = voice_name_B
if not text_to_speech(line, temp_file, voice_name=voice_name):
print(f"生成 {filename} 的第 {idx+1} 行音频失败")
# 清理临时文件
for temp_file_to_remove in temp_files:
if os.path.exists(temp_file_to_remove):
os.remove(temp_file_to_remove)
return
if not temp_files:
print(f"没有为 {filename} 生成音频")
return
# 使用 ffmpeg 拼接
concat_file = os.path.join(OUTPUT_DIRECTORY, "concat.txt")
with open(concat_file, 'w') as f:
for temp_file in temp_files:
f.write(f"file '{os.path.abspath(temp_file)}'\n")
try:
subprocess.run(
['ffmpeg', '-f', 'concat', '-safe', '0', '-i', concat_file, '-c', 'copy', output_filename],
check=True,
capture_output=True
)
print(f"成功拼接音频到 {output_filename}")
except subprocess.CalledProcessError as e:
print(f"拼接音频时出错：{e.stderr.decode()}")
finally:
os.remove(concat_file)
for temp_file in temp_files:
os.remove(temp_file)
if __name__ == "__main__":
parser = argparse.ArgumentParser(description="处理对话JSON文件以生成音频。")
args = parser.parse_args()
os.makedirs(OUTPUT_DIRECTORY, exist_ok=True)
for filename in os.listdir(INPUT_DIRECTORY):
if filename.endswith(".json"):
process_conversation(filename)
封面
ffmpeg -i deepseek.jpg -vf "crop=854:480" deepseek_480p_cropped.jpg
视频
ffmpeg -loop 1 -i deepseek.jpg -i deepseek.mp3 -c:v libx264 -tune stillimage -c:a aac -b:a 192k -shortest output_video.mp4