大家好，我平时喜欢在周末去公园散步，顺便买一杯三十五块钱的手冲咖啡。
你要问我为什么喜欢这样安排？其实很简单，安静的早晨能让人想清楚很多事情。
今天是七月三十号，天气不错，希望你听到这段声音的时候，心情也一样好。

$ curl -s https://api.minimaxi.com/v1/files/upload \
    -H "Authorization: Bearer $MINIMAX_KEY" \
    -F purpose=voice_clone \
    -F file=@sample.m4a
# 返回（记下 file_id 这串数字）：
{"file":{"file_id":312690747959545,"bytes":512340,"purpose":"voice_clone",...},
 "base_resp":{"status_code":0,"status_msg":"success"}}

$ curl -s https://api.minimaxi.com/v1/voice_clone \
    -H "Authorization: Bearer $MINIMAX_KEY" \
    -H "Content-Type: application/json" \
    -d '{
      "file_id": 312690747959545,
      "voice_id": "MyVoice20260730",
      "text": "你好，这是我的克隆音色试听，听听像不像我本人。",
      "model": "speech-2.8-hd",
      "need_noise_reduction": true,
      "need_volume_normalization": true
    }'
# 返回里 demo_audio 是试听音频链接，浏览器打开直接听：
{"input_sensitive":false,
 "demo_audio":"https://.../demo_xxx.mp3",
 "extra_info":{"audio_length":4830,"usage_characters":22,...},
 "base_resp":{"status_code":0,"status_msg":"success"}}

$ curl -s https://api.minimaxi.com/v1/t2a_v2 \
    -H "Authorization: Bearer $MINIMAX_KEY" \
    -H "Content-Type: application/json" \
    -d '{
      "model": "speech-2.8-hd",
      "text": "您好，我是您的专属助理，本月账单已经生成，共消费三百二十一元。",
      "output_format": "url",
      "voice_setting": { "voice_id": "MyVoice20260730", "speed": 1.0 }
    }'
# 返回 data.audio 是 mp3 链接（24 小时有效），下载保存：
{"data":{"audio":"https://.../xxx.mp3",...},
 "extra_info":{"usage_characters":30,...},
 "base_resp":{"status_code":0}}

import os, requests

KEY  = os.environ["MINIMAX_KEY"]          # export MINIMAX_KEY=xxx
BASE = "https://api.minimaxi.com"
H    = {"Authorization": f"Bearer {KEY}"}

# ① 上传样本 → file_id
up = requests.post(f"{BASE}/v1/files/upload", headers=H,
        data={"purpose": "voice_clone"},
        files={"file": open("sample.m4a", "rb")}).json()
file_id = up["file"]["file_id"]
print("file_id =", file_id)

# ② 复刻 → voice_id（带试听）
clone = requests.post(f"{BASE}/v1/voice_clone", headers=H, json={
    "file_id": file_id,
    "voice_id": "MyVoice20260730",
    "text": "你好，这是我的克隆音色试听。",
    "model": "speech-2.8-hd",
    "need_noise_reduction": True,
    "need_volume_normalization": True,
}).json()
print("试听链接 =", clone.get("demo_audio"))

# ③ 正式合成 → 下载 mp3（首次调用扣 9.9 元音色费）
tts = requests.post(f"{BASE}/v1/t2a_v2", headers=H, json={
    "model": "speech-2.8-hd",
    "text": "您好，我是您的专属助理，本月账单已经生成。",
    "output_format": "url",
    "voice_setting": {"voice_id": "MyVoice20260730", "speed": 1.0},
}).json()
url = tts["data"]["audio"]
open("cloned_output.mp3", "wb").write(requests.get(url).content)
print("已保存 cloned_output.mp3")

$ pip3 install openai-whisper jiwer
$ python3 - <<'EOF'
import whisper, jiwer
ref = "您好我是您的专属助理本月账单已经生成"   # 合成时用的原文（去标点）
m = whisper.load_model("small")
hyp = m.transcribe("cloned_output.mp3", language="zh")["text"]
hyp = "".join(c for c in hyp if c.strip() and c not in "，。！？、")
print("转写:", hyp)
print("CER = %.1f%%" % (jiwer.cer(ref, hyp) * 100))   # 中文用字错误率 CER
EOF
转写: 您好我是您的专属助理本月账单已经生成
CER = 0.0%   # 经验线：<3% 可用，<1.5% 优秀，>8% 说明合成含混

$ pip3 install speechbrain torchaudio
$ python3 - <<'EOF'
from speechbrain.inference.speaker import SpeakerRecognition
sv = SpeakerRecognition.from_hparams("speechbrain/spkrec-ecapa-voxceleb",
                                     savedir="./sv_model")
score, same = sv.verify_files("real_new.wav", "cloned_output.mp3")
print("SIM = %.3f  同一人判定 = %s" % (score, bool(same)))
EOF
SIM = 0.712  同一人判定 = True   # 经验线：>0.5 判同人，>0.7 相似度优秀