第30课：毕业项目：智能语音助手

import numpy as np class SimpleKWS: def detect(self, audio): return np.mean(audio**2) > 0.1 class SimpleASR: def recognize(self, mel): results = ["今天天气怎么样","你好","设置闹钟","播放音乐","打开灯","现在几点了"] return results[np.random.randint(len(results))] class SimpleNLU: def __init__(self): self.intents = {'weather':['天气','温度'],'greeting':['你好','嗨'],'alarm':['闹钟','提醒'], 'music':['播放','音乐'],'control':['打开','关闭'],'time':['几点','时间']} def parse(self, text): for intent, kws in self.intents.items(): for kw in kws: if kw in text: return {'intent':intent,'confidence':0.9,'slots':{}} return {'intent':'unknown','confidence':0.0,'slots':{}} class DialogueManager: def __init__(self): self.responses = {'weather':'今天晴天，气温25度。','greeting':'你好！有什么可以帮你的？', 'alarm':'好的，已设置闹钟。','music':'正在播放音乐🎵','control':'已执行操作。', 'time':'现在是下午3点30分。','unknown':'抱歉，请再说一遍。'} def respond(self, nlu): return self.responses.get(nlu['intent'], self.responses['unknown']) class SimpleTTS: def synthesize(self, text): sr = 16000; dur = len(text)*0.15 t = np.linspace(0, dur, int(sr*dur), endpoint=False) sig = np.zeros_like(t) for h in range(1,8): sig += (0.6**h)*np.sin(2*np.pi*120*h*t) return sig/(np.max(np.abs(sig))+1e-10)*0.5, sr class VoiceAssistant: def __init__(self): self.kws, self.asr, self.nlu, self.dm, self.tts = SimpleKWS(), SimpleASR(), SimpleNLU(), DialogueManager(), SimpleTTS() self.listening = False; self.history = [] def process(self, audio, mel=None): if not self.listening: if self.kws.detect(audio): self.listening = True; return "🎙️ 正在听..." return None text = self.asr.recognize(mel) nlu = self.nlu.parse(text) resp = self.dm.respond(nlu) audio_out, sr = self.tts.synthesize(resp) self.listening = False; self.history.append({'user':text,'response':resp}) return f"用户: {text}\n助手: {resp}" asst = VoiceAssistant() for audio in [np.random.randn(16000)*0.001, np.random.randn(16000)*0.3, np.random.randn(48000)*0.2]: r = asst.process(audio) if r: print(r) print("✅ 智能语音助手验证通过")

# 30课学习路线回顾 print('='*50) print('📡 信号处理(1-6): 波形/采样/FFT/STFT/MFCC/IO') print('🧠 声学模型(7-12): 特征/GMM-HMM/DNN-HMM/CTC/RNN-T/注意力') print('🚀 端到端ASR(13-18): LAS/Transformer/Conformer/Whisper/LM/解码') print('🗣️ 语音合成(19-24): TTS基础/前端/Tacotron/声码器/VITS/克隆') print('⚡ 实战项目(25-30): KWS/说话人/增强/多语言/流式/助手') print('='*50) print("✅ 全部30课总结完成")

import numpy as np import torch import torch.nn as nn import time np.random.seed(42) print("===== 参数扫描实验 =====") # 测试不同模型配置 configs = [ {'name': 'small', 'hidden': 64, 'layers': 2}, {'name': 'medium', 'hidden': 128, 'layers': 3}, {'name': 'large', 'hidden': 256, 'layers': 4}, {'name': 'xlarge', 'hidden': 512, 'layers': 6}, ] results = [] for cfg in configs: layers = [] for i in range(cfg['layers']): in_d = 40 if i == 0 else cfg['hidden'] layers.extend([nn.Linear(in_d, cfg['hidden']), nn.ReLU(), nn.Dropout(0.1)]) layers.append(nn.Linear(cfg['hidden'], 10)) model = nn.Sequential(*layers) params = sum(p.numel() for p in model.parameters()) x = torch.randn(16, 40) # 前向传播计时 start = time.time() for _ in range(100): _ = model(x) elapsed = time.time() - start results.append({'name': cfg['name'], 'params': params, 'time_ms': elapsed*10}) print(f" {cfg['name']:8s}: {params:>8,} params, {elapsed*10:.2f}ms/iter") print(" 结论: 模型越大越准但越慢，需在精度和速度间取舍") print("✅ 参数扫描实验完成")

import numpy as np import torch import torch.nn as nn import torch.nn.functional as F print('='*60) print('第30课 综合实验') print('='*60) sr = 16000 t = np.linspace(0, 1, sr, endpoint=False) np.random.seed(42) test_sig = 0.8*np.sin(2*np.pi*440*t) + 0.3*np.sin(2*np.pi*1000*t) + 0.01*np.random.randn(sr) print('信号统计: RMS=%.4f, Peak=%.4f' % (np.sqrt(np.mean(test_sig**2)), np.max(np.abs(test_sig)))) fft_mag = np.abs(np.fft.rfft(test_sig)) fft_freqs = np.fft.rfftfreq(sr, 1/sr) top_freqs = fft_freqs[np.argsort(fft_mag)[-5:]] print('主要频率成分:', sorted(top_freqs.astype(int)), 'Hz') class TestModel(nn.Module): def __init__(self, input_dim=80, hidden=128, output=10): super().__init__() self.encoder = nn.LSTM(input_dim, hidden, 2, batch_first=True, bidirectional=True) self.classifier = nn.Linear(hidden*2, output) def forward(self, x): out, _ = self.encoder(x) return self.classifier(out[:, -1, :]) model = TestModel() x = torch.randn(4, 50, 80) out = model(x) print('模型: 输入', x.shape, '-> 输出', out.shape) print('参数量:', sum(p.numel() for p in model.parameters())) optimizer = torch.optim.Adam(model.parameters(), lr=1e-3) criterion = nn.CrossEntropyLoss() for step in range(5): x = torch.randn(8, 50, 80) y = torch.randint(0, 10, (8,)) out = model(x) loss = criterion(out, y) optimizer.zero_grad() loss.backward() optimizer.step() print('训练5步后loss: %.4f' % loss.item()) model.eval() with torch.no_grad(): correct = 0; total = 0 for _ in range(10): x = torch.randn(16, 50, 80) y = torch.randint(0, 10, (16,)) pred = model(x).argmax(1) correct += (pred == y).sum().item() total += 16 print('随机初始化模型准确率: %.1f%% (期望约10%%)' % (correct/total*100)) print('✅ 综合实验验证通过')