esp32s3聊天机器人（二）

继续上文，硬件软件准备齐全，介绍一下主要用到的库

sherpa-onnx

开源的，语音转文本、文本转语音、说话人分类和 VAD，关键是支持C#开发

OllamaSharp

用于连接ollama，如其名C#开发

虽然离可玩还有一段距离，但是还是要说一下目前遇到的一些问题

1、最初使用流式语音识别，但录音流发送到server之后，完全识别不到文字，保存为pcm播放发现完全是噪音，最后转换放大了音频可以正确识别到说话内容了

static float gain = 5.0f;
short[] int16Array;
float[] floatArray;
public void Recognize(byte[] bytes)
{
    //Console.WriteLine("收到音频长度："+ bytes.Length);
    int16Array = new short[bytes.Length / 2];
    Buffer.BlockCopy(bytes, 0, int16Array, 0, bytes.Length);
    floatArray = new float[int16Array.Length];
    for (int i = 0; i < int16Array.Length; i++)
    {
        floatArray[i] = int16Array[i] / 32768.0f * gain;
    }
    onlineStream.AcceptWaveform(sampleRate, floatArray);
}

2、流式识别有个问题，我说下一句话的时候才会确定上一句结束，虽然有三个参数可调，但我多次测试，没调好，只好改用离线识别，给esp32s3添加个按键，按下开始发送录音数据，松开停止发送并告知录音结束，server端再开始识别

            //默认值
            config.Rule1MinTrailingSilence = 2.4f;
            config.Rule2MinTrailingSilence = 0.5f;
            //限制最长说话10秒
            config.Rule3MinUtteranceLength = 10f;

void loop() {
  webSocket.loop(); // 必须调用以处理WebSocket事件 
  if(digitalRead(PIN_BUTTON) == LOW)
  {
    pressed = true;
    uint8_t buffer[BUFFER_SIZE];
    size_t bytesRead;
    // 从I2S读取音频数据
    i2s_read(I2S_NUM_0, buffer, BUFFER_SIZE, &bytesRead, portMAX_DELAY);
    // 通过WebSocket发送音频数据
    if (webSocket.sendBIN(buffer, bytesRead)) {
      //Serial.printf("Sent %d bytes of audio data\n", bytesRead); 
    } else {
      //Serial.println("Failed to send audio data");
    }
  }
  else
  {
    if(pressed)
    { 
      pressed = false;
      if (webSocket.sendTXT("{\"code\":1,\"message\":\"结束语音\"}")) {
       
    } else {
       
    }
    }
  }
}

        private static void OnMessage(IWebSocketConnection connection, string msg)
        {
            BaseMsg baseMsg = null;
            try
            {
                baseMsg = JsonConvert.DeserializeObject<BaseMsg>(msg);
            }
            catch (Exception e)
            {
                Console.WriteLine(e);
            }
            if (baseMsg != null)
            {
                // 收到code 1时，结束录音开始识别
                if (baseMsg.code == 1)
                {
                    Asr asr = null;
                    asrs.TryGetValue(connection.GetHashCode(), out asr);
                    if (asr != null)
                    {
                        asr.EndReceive();
                    }
                }
            }
        }

3、周围人多声音嘈杂的时候，语音识别开始放飞自我，七八米开外的人声它都收进来…… 用RNNoise.Net处理了一下，也不知道对不对，感觉有点效果……

https://github.com/Yellow-Dog-Man/RNNoise.Net

        /// <summary>
        /// 识别语音数据
        /// </summary>
        short[] int16Array;
        float[] floatArray;
        private void Recognize(byte[] bytes)
        {
            int16Array = new short[bytes.Length / 2];
            Buffer.BlockCopy(bytes, 0, int16Array, 0, bytes.Length);
            floatArray = new float[int16Array.Length];
            for (int i = 0; i < int16Array.Length; i++)
            {
                floatArray[i] = int16Array[i] / 32768.0f;
            }
            // 降噪
            using (var denoiser = new Denoiser())
            {
                int count = denoiser.Denoise(floatArray.AsSpan());
                Console.WriteLine("denoised count:" + count);
            }

            offlineStream = recognizer.CreateStream();
            offlineStream.AcceptWaveform(sampleRate, floatArray);
            recognizer.Decode(offlineStream);
            string result = offlineStream.Result.Text;
            offlineStream.Dispose();
            Console.WriteLine("result:" + result);
            if (!string.IsNullOrWhiteSpace(result))
            {
                result = offlinePunctuation.AddPunct(result.ToLower());
                BaseMsg textMsg = new BaseMsg(1, result);
                client.Send(JsonConvert.SerializeObject(textMsg));
                if (llm != null)
                {
                    llm.RequestAsync(result);
                }
            }
        }