Question

因此，我使用keras创建了一个模型，该模型将声音分类为某些单词。我用来创建模型的数据集只有1秒钟的声音。我想用C＃或Python创建一个应用程序，它将使用我创建的模型从麦克风听一些单词（不是从文件输入声音数据，而是从麦克风实时输入）。如果我说一些话，那么应用程序将执行某些操作。

这是我在Python中的尝试，但效果不佳，因为预测始终对错误的单词有100％的把握。

import collections
import contextlib
import functools
import queue
import pyaudio
from ctypes import windll
from keras.models import load_model
import numpy as np

model = load_model("..\\model.hdf5")

labels = ["roll left", "roll right", "yaw left", "yaw right", "pitch up", "pitch down", "stop"]


def read(data):
    probability = model.predict(data.reshape(1, 8000, 1))
    index = np.argmax(probability[0])
    print(probability)
    print(labels[index], probability[0][index])
    # if word == something
    #     do some job


class SpeechRecog:
    CHANNELS = 1
    RATE = 8000
    CHUNK = RATE
    SECS_OVERLAP = 1

    def __init__(self):
        self.audio_interface = pyaudio.PyAudio()

    def listen(self):
        with self.record_audio() as buff:
            overlap_buffer = collections.deque(maxlen=int(self.SECS_OVERLAP * self.RATE / self.CHUNK))
            for data in self._audio_data_generator(buff, overlap_buffer):
                read(np.frombuffer(data, dtype=np.int16))

    @staticmethod
    def _audio_data_generator(buff: queue.Queue, overlap_buffer):
        if overlap_buffer:
            yield b''.join(overlap_buffer)
            overlap_buffer.clear()

        while True:
            # Use a blocking get() to ensure there's at least one chunk of data.
            data = [buff.get()]

            # Now consume whatever other data's still buffered.
            while True:
                try:
                    data.append(buff.get(block=False))
                except queue.Empty:
                    break

            # `None` in the buffer signals that we should stop generating. Put the
            # data back into the buffer for the next generator.
            if None in data:
                data.remove(None)
                if data:
                    buff.put(b''.join(data))
                break
            else:
                overlap_buffer.extend(data)

            yield b''.join(data)

    def fill_buffer(self, buff: queue.Queue, in_data, frame_count, time_info, status_flag):
        buff.put(in_data)
        return None, pyaudio.paContinue

    @contextlib.contextmanager
    def record_audio(self):
        buff: queue.Queue = queue.Queue()
        stream: pyaudio.Stream = self.audio_interface \
            .open(format=pyaudio.paInt16,
                  channels=self.CHANNELS,
                  rate=self.RATE,
                  input=True,
                  frames_per_buffer=self.CHUNK,
                  stream_callback=functools.partial(self.fill_buffer, buff))

        yield buff

        stream.stop_stream()
        stream.close()

        self.audio_interface.terminate()


s = SpeechRecog()
s.listen()

您能给我一些如何解决此问题的指示吗？

从训练有素的模型创建实时语音识别

0 个答案: