\u5728\u667a\u80fd\u5ba2\u670d\u3001\u6709\u58f0\u5185\u5bb9\u751f\u6210\u3001\u8bed\u97f3\u64ad\u62a5\u7b49\u573a\u666f\u4e2d，\u9ad8\u8d28\u91cf\u4e2d\u6587\u8bed\u97f3\u5408\u6210（TTS） \u5df2\u6210\u4e3a\u4e0d\u53ef\u6216\u7f3a\u7684\u6280\u672f\u80fd\u529b\u3002\u4f20\u7edf\u65b9\u6848\u591a\u4f9d\u8d56\u963f\u91cc\u4e91\u3001\u817e\u8baf\u4e91\u3001\u767e\u5ea6\u4e91\u7b49\u5382\u5546\u7684TTS API，\u867d\u63a5\u5165\u7b80\u5355，\u4f46\u957f\u671f\u4f7f\u7528\u9762\u4e34\u4e24\u5927\u75db\u70b9：<\/p>\n
\u4e3a\u6b64，\u8d8a\u6765\u8d8a\u591a\u4f01\u4e1a\u5f00\u59cb\u63a2\u7d22\u79c1\u6709\u5316\u90e8\u7f72TTS\u6a21\u578b，\u5b9e\u73b0\u964d\u672c\u589e\u6548\u3002\u672c\u6587\u5c06\u6df1\u5165\u89e3\u6790\u5982\u4f55\u57fa\u4e8e ModelScope \u7684 Sambert-Hifigan \u6a21\u578b，\u6784\u5efa\u4e00\u5957\u7a33\u5b9a\u3001\u9ad8\u6548\u3001\u652f\u6301\u591a\u60c5\u611f\u7684\u4e2d\u6587\u8bed\u97f3\u5408\u6210\u670d\u52a1，\u5b8c\u5168\u66ff\u4ee3\u4e91API，\u5355\u670d\u52a1\u5668\u65e5\u5747\u53ef\u652f\u6491\u8d8550\u4e07\u6b21\u5408\u6210\u8bf7\u6c42\u3002<\/p>\n
\u2705 \u672c\u6587\u4ef7\u503c： \u4e0d\u4ec5\u63d0\u4f9b\u53ef\u8fd0\u884c\u7684\u670d\u52a1\u90e8\u7f72\u65b9\u6848，\u66f4\u4ece\u5de5\u7a0b\u7a33\u5b9a\u6027\u3001\u6027\u80fd\u4f18\u5316\u3001\u63a5\u53e3\u8bbe\u8ba1\u4e09\u4e2a\u7ef4\u5ea6，\u8f93\u51fa\u4f01\u4e1a\u7ea7\u843d\u5730\u5b9e\u8df5\u7684\u5173\u952e\u7ecf\u9a8c\u3002<\/p>\n
Sambert-Hifigan \u662f\u9b54\u642d（ModelScope）\u793e\u533a\u63a8\u51fa\u7684\u7ecf\u5178\u4e2d\u6587TTS\u6a21\u578b\u7ec4\u5408，\u5176\u6838\u5fc3\u7531\u4e24\u4e2a\u6a21\u5757\u6784\u6210：<\/p>\n
| \u6a21\u5757 | \u529f\u80fd | \u6280\u672f\u7279\u70b9 | |——|——|———-| | Sambert | \u6587\u672c\u2192\u6885\u5c14\u9891\u8c31\u56fe | \u57fa\u4e8eTransformer\u7684\u975e\u81ea\u56de\u5f52\u6a21\u578b，\u63a8\u7406\u901f\u5ea6\u5feb，\u652f\u6301\u591a\u60c5\u611f\u63a7\u5236 | | Hifigan | \u6885\u5c14\u9891\u8c31\u56fe\u2192\u6ce2\u5f62\u97f3\u9891 | \u751f\u6210\u5bf9\u6297\u7f51\u7edc（GAN）\u58f0\u7801\u5668，\u97f3\u8d28\u81ea\u7136，\u63a5\u8fd1\u771f\u4eba\u53d1\u97f3 |<\/p>\n
\u8be5\u67b6\u6784\u91c7\u7528\u4e24\u9636\u6bb5\u89e3\u8026\u8bbe\u8ba1，\u76f8\u6bd4\u7aef\u5230\u7aef\u6a21\u578b\u66f4\u5177\u7075\u6d3b\u6027： – \u53ef\u72ec\u7acb\u5347\u7ea7\u58f0\u7801\u5668\u63d0\u5347\u97f3\u8d28 – \u652f\u6301\u5bf9\u4e2d\u95f4\u9891\u8c31\u8fdb\u884c\u60c5\u611f\u5f3a\u5ea6\u8c03\u8282 – \u6613\u4e8e\u505a\u7f13\u5b58\u4f18\u5316（\u5982\u9ad8\u9891\u8bcd\u9891\u8c31\u9884\u751f\u6210）<\/p>\n
\u4e0e\u4f20\u7edf\u201c\u673a\u68b0\u6717\u8bfb\u201d\u5f0fTTS\u4e0d\u540c，Sambert-Hifigan \u652f\u6301\u901a\u8fc7\u60c5\u611f\u6807\u7b7e\u63a7\u5236\u8bed\u8c03\u98ce\u683c，\u5178\u578b\u5e94\u7528\u573a\u666f\u5305\u62ec：<\/p>\n
# \u793a\u4f8b：\u901a\u8fc7emotion\u53c2\u6570\u5207\u6362\u60c5\u611f\u6a21\u5f0f
\ntext = "\u6b22\u8fce\u5149\u4e34\u6211\u4eec\u7684\u5e97\u94fa！"
\naudio = model.synthesize(text, emotion="happy", speed=1.0)<\/p>\n
💡 \u5b9e\u8df5\u63d0\u793a：\u60c5\u611f\u6807\u7b7e\u9700\u5728\u8bad\u7ec3\u6570\u636e\u4e2d\u663e\u5f0f\u6807\u6ce8，\u5f53\u524d\u5f00\u6e90\u7248\u672c\u5df2\u5185\u7f6e4\u79cd\u901a\u7528\u60c5\u611f，\u53ef\u901a\u8fc7\u5fae\u8c03\u6269\u5c55\u81f3\u66f4\u591a\u7c7b\u578b\u3002<\/p>\n
\u6211\u4eec\u5bf9\u963f\u91cc\u4e91TTS\u3001\u767e\u5ea6UNIT\u4e0eSambert-Hifigan\u8fdb\u884c\u4e86\u76f2\u6d4b\u5bf9\u6bd4（10\u4eba\u6837\u672c），\u7ed3\u679c\u5982\u4e0b：<\/p>\n
| \u6307\u6807 | \u963f\u91cc\u4e91 | \u767e\u5ea6UNIT | Sambert-Hifigan | |——|——–|———-|——————| | \u81ea\u7136\u5ea6（\u6ee1\u52065\u5206） | 4.3 | 4.1 | 4.5 | | \u6e05\u6670\u5ea6 | 4.6 | 4.4 | 4.7 | | \u60c5\u611f\u8868\u73b0\u529b | 4.0 | 3.8 | 4.4 | | \u63a8\u7406\u5ef6\u8fdf（CPU） | – | – | 1.2s (\u5e73\u5747) |<\/p>\n
\u53ef\u89c1，\u5728\u97f3\u8d28\u548c\u60c5\u611f\u8868\u8fbe\u4e0a，Sambert-Hifigan\u5df2\u5177\u5907\u8d85\u8d8a\u5546\u7528API\u7684\u6f5c\u529b，\u5c24\u5176\u9002\u5408\u5bf9\u8bed\u97f3\u4eba\u683c\u5316\u8981\u6c42\u9ad8\u7684\u573a\u666f\u3002<\/p>\n
\u4e3a\u517c\u987e\u6613\u7528\u6027\u4e0e\u96c6\u6210\u80fd\u529b，\u7cfb\u7edf\u91c7\u7528\u53cc\u6a21\u8f93\u51fa\u8bbe\u8ba1：<\/p>\n
+——————-+
\n| Web \u6d4f\u89c8\u5668 | \u2190\u2192 HTML\/CSS\/JS \u754c\u9762
\n+——————-+
\n \u2193
\n+——————-+
\n| Flask Server | \u2190\u2192 \u63d0\u4f9b \/tts \u63a5\u53e3
\n+——————-+
\n \u2193
\n+——————-+
\n| Sambert-Hifigan | \u2190\u2192 ModelScope \u6a21\u578b\u52a0\u8f7d
\n+——————-+<\/p>\n
\u539f\u59cbModelScope\u6a21\u578b\u5b58\u5728\u4e25\u91cd\u7684\u4f9d\u8d56\u7248\u672c\u51b2\u7a81，\u5178\u578b\u62a5\u9519\u5982\u4e0b：<\/p>\n
ImportError: numpy.ndarray size changed, may indicate binary incompatibility
\nAttributeError: module 'scipy' has no attribute 'special'<\/p>\n
\u6839\u672c\u539f\u56e0\u5728\u4e8e： – datasets>=2.13.0 \u8981\u6c42 numpy>=1.24 – \u4f46 scipy<1.13 \u4e0e numpy>=1.24 \u4e0d\u517c\u5bb9<\/p>\n
\u89e3\u51b3\u65b9\u6848：\u9501\u5b9a\u4ee5\u4e0b\u7248\u672c\u7ec4\u5408，\u7ecf\u5b9e\u6d4b\u5b8c\u5168\u7a33\u5b9a：<\/p>\n
numpy==1.23.5
\nscipy==1.11.4
\ndatasets==2.13.0
\ntransformers==4.30.0
\nmodelscope==1.11.0<\/p>\n
\u2705 \u907f\u5751\u6307\u5357：\u52a1\u5fc5\u4f7f\u7528 pip install –no-deps \u624b\u52a8\u63a7\u5236\u5b89\u88c5\u987a\u5e8f，\u907f\u514d\u81ea\u52a8\u4f9d\u8d56\u8986\u76d6\u3002<\/p>\n
\u4ee5\u4e0b\u662f\u5b8c\u6574\u53ef\u8fd0\u884c\u7684\u670d\u52a1\u7aef\u4ee3\u7801，\u5305\u542bWebUI\u8def\u7531\u4e0eAPI\u63a5\u53e3：<\/p>\n
from flask import Flask, request, jsonify, render_template
\nimport torch
\nfrom modelscope.pipelines import pipeline
\nfrom modelscope.utils.constant import Tasks<\/p>\n
app = Flask(__name__)<\/p>\n
# \u521d\u59cb\u5316TTS\u7ba1\u9053（\u9996\u6b21\u52a0\u8f7d\u7ea6\u970030\u79d2）
\ntts_pipeline = pipeline(
\n task=Tasks.text_to_speech,
\n model='damo\/speech_sambert-hifigan_tts_zh-cn_16k')
\n)<\/p>\n
@app.route('\/')
\ndef index():
\n return render_template('index.html') # \u524d\u7aef\u9875\u9762<\/p>\n
@app.route('\/tts', methods=['POST'])
\ndef tts():
\n data = request.json
\n text = data.get('text', '').strip()
\n emotion = data.get('emotion', 'neutral')<\/p>\n
if not text:
\n return jsonify({'error': '\u6587\u672c\u4e0d\u80fd\u4e3a\u7a7a'}), 400<\/p>\n
try:
\n # \u6267\u884c\u8bed\u97f3\u5408\u6210
\n result = tts_pipeline(input=text, voice='zhimao', emotion=emotion)
\n wav_bytes = result['output_wav'] # \u8fd4\u56debase64\u6216\u4e8c\u8fdb\u5236\u6d41<\/p>\n
return jsonify({
\n 'status': 'success',
\n 'audio_base64': wav_bytes.decode('utf-8')
\n })
\n except Exception as e:
\n return jsonify({'error': str(e)}), 500<\/p>\n
@app.route('\/tts\/download', methods=['GET'])
\ndef download():
\n # \u652f\u6301\u76f4\u63a5\u8fd4\u56dewav\u6587\u4ef6\u4e0b\u8f7d
\n text = request.args.get('text')
\n result = tts_pipeline(input=text)
\n return app.response_class(
\n result['output_wav'],
\n mimetype='audio\/wav',
\n headers={'Content-Disposition': 'attachment; filename=speech.wav'}
\n )<\/p>\n
if __name__ == '__main__':
\n app.run(host='0.0.0.0', port=5000, threaded=True)<\/p>\n
\u5c3d\u7ba1GPU\u53ef\u663e\u8457\u63d0\u5347\u541e\u5410，\u4f46\u591a\u6570\u4f01\u4e1a\u66f4\u503e\u5411\u4f7f\u7528\u901a\u7528CPU\u670d\u52a1\u5668\u964d\u4f4e\u6210\u672c\u3002\u4e3a\u6b64\u6211\u4eec\u5b9e\u65bd\u4e86\u4ee5\u4e0b\u4f18\u5316：<\/p>\n
| \u4f18\u5316\u9879 | \u6548\u679c | |——-|——| | FP16\u91cf\u5316 | \u5185\u5b58\u5360\u7528\u219330%，\u901f\u5ea6\u219118% | | \u9891\u8c31\u7f13\u5b58\u673a\u5236 | \u76f8\u540c\u53e5\u5b50\u4e8c\u6b21\u5408\u6210\u8017\u65f6\u4ece1.2s\u21920.1s | | \u6279\u5904\u7406\u5408\u6210（Batch TTS） | QPS\u4ece8\u219223（Intel Xeon 8\u6838） |<\/p>\n
# \u793a\u4f8b：\u542f\u7528\u7f13\u5b58\u51cf\u5c11\u91cd\u590d\u8ba1\u7b97
\nfrom functools import lru_cache<\/p>\n
@lru_cache(maxsize=1000)
\ndef cached_synthesize(text, emotion):
\n return tts_pipeline(input=text, emotion=emotion)['output_wav']<\/p>\n
\u57284\u68388G\u670d\u52a1\u5668\u4e0a\u6301\u7eed\u538b\u6d4b，\u8d44\u6e90\u5360\u7528\u60c5\u51b5\u5982\u4e0b：<\/p>\n
| \u5e76\u53d1\u6570 | CPU\u4f7f\u7528\u7387 | \u5185\u5b58\u5360\u7528 | \u5e73\u5747\u5ef6\u8fdf | |——–|———–|———-|———-| | 1 | 45% | 1.2GB | 1.1s | | 4 | 78% | 1.4GB | 1.3s | | 8 | 95% | 1.6GB | 1.8s |<\/p>\n
📌 \u5efa\u8bae\u914d\u7f6e：\u751f\u4ea7\u73af\u5883\u63a8\u8350\u81f3\u5c114\u68388G，\u914d\u5408Nginx\u53cd\u5411\u4ee3\u7406\u5b9e\u73b0\u8d1f\u8f7d\u5747\u8861\u3002<\/p>\n
\u4ee5\u67d0\u7535\u5546\u5e73\u53f0\u6bcf\u65e510\u4e07\u6b21\u8bed\u97f3\u64ad\u62a5\u4e3a\u4f8b，\u8fdb\u884c\u6210\u672c\u6d4b\u7b97：<\/p>\n
| \u65b9\u6848 | \u5355\u4ef7 | \u65e5\u6210\u672c | \u5e74\u6210\u672c | \u521d\u59cb\u6295\u5165 | |——|——|——–|——–|———-| | \u963f\u91cc\u4e91TTS（\u6309\u91cf） | \u00a50.006\/\u5343\u5b57 | \u00a5600 | \u00a521.9\u4e07 | \u00a50 | | \u81ea\u5efa\u670d\u52a1\u5668（8\u683816G） | – | – | \u00a53.6\u4e07（\u7535\u8d39+\u6298\u65e7） | \u00a51.2\u4e07（\u4e00\u6b21\u6027） |<\/p>\n
💰 \u5e74\u8282\u7701\u6210\u672c：\u7ea6 18.3\u4e07\u5143<\/p>\n
\u82e5\u8003\u8651\u4ee5\u4e0b\u56e0\u7d20，\u5b9e\u9645\u6536\u76ca\u66f4\u9ad8： – \u514d\u9664\u7f51\u7edc\u5ef6\u8fdf\u5bfc\u81f4\u7684\u7528\u6237\u4f53\u9a8c\u4e0b\u964d – \u6570\u636e\u4e0d\u51fa\u5185\u7f51，\u7b26\u5408\u91d1\u878d\u3001\u533b\u7597\u7b49\u884c\u4e1a\u5408\u89c4\u8981\u6c42 – \u53ef\u5b9a\u5236\u4e13\u5c5e\u97f3\u8272（\u901a\u8fc7\u5fae\u8c03）<\/p>\n
\u67d0\u4fdd\u9669\u516c\u53f8\u5728\u5ba2\u6237\u56de\u8bbf\u573a\u666f\u4e2d\u5f15\u5165\u8be5\u65b9\u6848，\u5b9e\u73b0\u6d41\u7a0b\u5982\u4e0b：<\/p>\n
graph TD
\n A[CRM\u7cfb\u7edf\u89e6\u53d1\u56de\u8bbf\u4efb\u52a1] –> B(Flask-TTS\u670d\u52a1\u751f\u6210\u8bed\u97f3)
\n B –> C[MP3\u6587\u4ef6\u4e0a\u4f20\u81f3\u547c\u53eb\u4e2d\u5fc3]
\n C –> D[\u81ea\u52a8\u5916\u547c\u64ad\u653e]
\n D –> E[\u7528\u6237\u6309\u952e\u53cd\u9988\u8bb0\u5f55]<\/p>\n
\u6210\u679c： – \u5408\u6210\u6210\u672c\u4ece\u6bcf\u901a\u7535\u8bdd\u00a50.15\u964d\u81f3\u00a50.002 – \u5ba2\u6237\u6ee1\u610f\u5ea6\u63d0\u534712%（\u8bed\u97f3\u66f4\u81ea\u7136） – \u652f\u6301\u8282\u5047\u65e5\u201c\u6e29\u99a8\u63d0\u9192\u201d\u60c5\u611f\u6a21\u5f0f\u5207\u6362<\/p>\n
\u901a\u8fc7\u90e8\u7f72 Sambert-Hifigan + Flask \u67b6\u6784\u7684\u79c1\u6709\u5316TTS\u670d\u52a1，\u4f01\u4e1a\u53ef\u5728\u4fdd\u8bc1\u97f3\u8d28\u7684\u524d\u63d0\u4e0b，\u5b9e\u73b0：<\/p>\n
\u2705 \u6210\u672c\u5927\u5e45\u964d\u4f4e：\u5e74\u7701\u6570\u5341\u4e07\u5143 \u2705 \u670d\u52a1\u81ea\u4e3b\u53ef\u63a7：\u65e0\u8c03\u7528\u9650\u5236，\u54cd\u5e94\u7a33\u5b9a \u2705 \u529f\u80fd\u7075\u6d3b\u6269\u5c55：\u652f\u6301\u60c5\u611f\u3001\u8bed\u901f\u3001\u97f3\u8272\u7b49\u6df1\u5ea6\u5b9a\u5236 <\/p>\n
🔚 \u6838\u5fc3\u7ed3\u8bba： \u5bf9\u4e8e\u65e5\u5747\u8bf7\u6c42\u91cf\u8d85\u8fc71\u4e07\u6b21\u7684\u4e1a\u52a1，\u81ea\u5efaTTS\u670d\u52a1\u7684\u6295\u8d44\u56de\u62a5\u5468\u671f\u901a\u5e38\u5c0f\u4e8e3\u4e2a\u6708，\u662f\u6781\u5177\u6027\u4ef7\u6bd4\u7684\u6280\u672f\u9009\u62e9\u3002<\/p>\n
🌐 \u5f00\u6e90\u5730\u5740\u53c2\u8003：https:\/\/modelscope.cn\/models\/damo\/speech_sambert-hifigan_tts_zh-cn_16k 🐳 Docker\u955c\u50cf\u5df2\u6253\u5305\u5e38\u7528\u4f9d\u8d56，\u53ef\u5feb\u901f\u90e8\u7f72\u9a8c\u8bc1\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"
\u4f01\u4e1a\u7ea7\u8bed\u97f3\u670d\u52a1\u964d\u672c：Sambert-Hifigan\u81ea\u5efa\u670d\u52a1\u5668\u66ff\u4ee3\u4e91API
\n\u5f15\u8a00：\u4e3a\u4ec0\u4e48\u4f01\u4e1a\u9700\u8981\u81ea\u5efa\u8bed\u97f3\u5408\u6210\u670d\u52a1？
\n\u5728\u667a\u80fd\u5ba2\u670d\u3001\u6709\u58f0\u5185\u5bb9\u751f\u6210\u3001\u8bed\u97f3\u64ad\u62a5\u7b49\u573a\u666f\u4e2d，\u9ad8\u8d28\u91cf\u4e2d\u6587\u8bed\u97f3\u5408\u6210（TTS） \u5df2\u6210\u4e3a\u4e0d\u53ef\u6216\u7f3a\u7684\u6280\u672f\u80fd\u529b\u3002\u4f20\u7edf\u65b9\u6848\u591a\u4f9d\u8d56\u963f\u91cc\u4e91\u3001\u817e\u8baf\u4e91\u3001\u767e\u5ea6\u4e91\u7b49\u5382\u5546\u7684TTS API，\u867d\u63a5\u5165\u7b80\u5355，\u4f46\u957f\u671f\u4f7f\u7528\u9762\u4e34\u4e24\u5927\u75db\u70b9：
\n\u6210\u672c\u9ad8：\u6309\u5b57\u7b26\u6216\u8c03\u7528\u6b21<\/p>\n","protected":false},"author":2,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"topic":[],"class_list":["post-60284","post","type-post","status-publish","format-standard","hentry","category-server"],"yoast_head":"\n