通过 Google Speech to Text 进行音频转文本输入

在本文中我们将研究以下主题

1. navigator.mediaDevices.getUserMedia 浏览器 Api
2. 谷歌语音到文本 API

我们将从创建 React hook 开始，它将执行所有操作，例如 startRecording、stopRecording、创建 Audio Blob、错误处理等。

在我们进入正题之前，还有一些其他事情需要注意

1. 最小分贝，高于该分贝我们将考虑将对话作为输入，例如-35db（只是一个随机数）
2. 表示用户已停止输入的暂停时间应该是多长，例如 2000 毫秒

const VOICE_MIN_DECIBELS = -35
const DELAY_BETWEEN_DIALOGUE = 2000

让我们将我们的钩子命名为useAudioInput.ts，我们将使用浏览器api，如navigator.mediaDevices.getUserMedia、MediaRecorder和AudioContext。 AudioContext 将帮助我们识别输入音频是否高于被视为输入所需的最小分贝，因此我们将从以下变量和道具开始

const defaultConfig = {
    audio: true
};

type Payload = Blob;

type Config = {
    audio: boolean;
    timeSlice?: number
    timeInMillisToStopRecording?: number
    onStop: () => void;
    onDataReceived: (payload: Payload) => void
};

export const useAudioInput = (config: Config = defaultConfig) => {
    const mediaChunks = useRef<Blob[]>([]);
    const [isRecording, setIsRecording] = useState(false);
    const mediaRecorder = useRef<MediaRecorder | null>(null);
    const [error, setError] = useState<Error| null>(null);
    let requestId: number;
    let timer: ReturnType<typeof setTimeout>;

    const createBlob = () => {
      const [chunk] = mediaChunks.current;
      const blobProperty = { type: chunk.type };
      return new Blob(mediaChunks.current, blobProperty)
    }
  ...
}

在上面的代码中，我们将使用 mediaChunks 作为变量来保存输入 blob 和 mediaRecorder 来拥有新 MediaRecorder 的实例，该实例将流作为来自 navigator.mediaDevices.getUserMedia 的输入。接下来让我们处理 getUserMedia 不可用的情况

...
useEffect(() => {
        if(!navigator.mediaDevices || !navigator.mediaDevices.getUserMedia) {
            const notAvailable = new Error('Your browser does not support Audio Input')
            setError(notAvailable)
        }

    },[]);
...

我们将开始编写钩子的实际功能，其中包括各种函数，如 setupMediaRecorder、setupAudioContext、onRecordingStart、onRecordingActive、startRecording、stopRecording 等。

const onRecordingStart = () => mediaChunks.current = [];

const onRecordingActive = useCallback(({data}: BlobEvent) => {
        if(data) {
            mediaChunks.current.push(data);
            config?.onDataReceived?.(createBlob())
        }
    },[config]);

const startTimer = () => {
        timer = setTimeout(() => {
            stopRecording();
        }, config.timeInMillisToStopRecording)
    };

const setupMediaRecorder = ({stream}:{stream: MediaStream}) => {
        mediaRecorder.current = new MediaRecorder(stream)
        mediaRecorder.current.ondataavailable = onRecordingActive
        mediaRecorder.current.onstop = onRecordingStop
        mediaRecorder.current.onstart = onRecordingStart
        mediaRecorder.current.start(config.timeSlice)

    };

 const setupAudioContext = ({stream}:{stream: MediaStream}) => {
        const audioContext = new AudioContext();
        const audioStreamSource = audioContext.createMediaStreamSource(stream);
        const analyser = audioContext.createAnalyser();

        analyser.minDecibels = VOICE_MIN_DECIBELS;

        audioStreamSource.connect(analyser);
        const bufferLength = analyser.frequencyBinCount;
        const domainData = new Uint8Array(bufferLength)

        return {
            domainData,
            bufferLength,
            analyser
        }
    };

const startRecording = async () => {
        setIsRecording(true);

        await navigator.mediaDevices
            .getUserMedia({
                audio: config.audio
            })
            .then((stream) => {
                setupMediaRecorder({stream});
                if(config.timeSlice) {
                    const { domainData, analyser, bufferLength } = setupAudioContext({ stream });
                    startTimer()
                }
            })
            .catch(e => {
                setError(e);
                setIsRecording(false)
            })
    };



    const stopRecording = () => {
        mediaRecorder.current?.stop();

        clearTimeout(timer);
        window.cancelAnimationFrame(requestId);

        setIsRecording(false);
        onRecordingStop()
    };

    const createBlob = () => {
        const [chunk] = mediaChunks.current;
        const blobProperty = { type: chunk.type };
        return new Blob(mediaChunks.current, blobProperty)
    }

    const onRecordingStop = () => config?.onStop?.();

通过上面的代码，我们已经差不多完成了钩子，唯一悬而未决的事情是识别用户是否已经停止说话，如果没有输入 2，我们将使用 DELAY_BETWEEN_DIALOGUE 作为等待的时间秒后，我们将假设用户已停止讲话并将点击语音转文本端点。

...
const detectSound = ({ 
        recording,
        analyser,
        bufferLength,
        domainData
    }: {
        recording: boolean
        analyser: AnalyserNode
        bufferLength: number
        domainData: Uint8Array
    }) => {
        let lastDetectedTime = performance.now();
        let anySoundDetected = false;

        const compute = () => {
            if (!recording) {
                return;
            }

            const currentTime = performance.now();

            const timeBetweenTwoDialog =
                anySoundDetected === true && currentTime - lastDetectedTime > DELAY_BETWEEN_DIALOGUE;

            if (timeBetweenTwoDialog) {
                stopRecording();

                return;
            }

            analyser.getByteFrequencyData(domainData);

            for (let i = 0; i < bufferLength; i += 1) {
                if (domainData[i] > 0) {
                    anySoundDetected = true;
                    lastDetectedTime = performance.now();
                }
            }

            requestId = window.requestAnimationFrame(compute);
        };

        compute();

    }
...

const startRecording = async () => {
 ... 
  detectSound()
 ... 
}

在上面的代码中，我们使用 requestAnimationFrame 来检测用户音频输入，这样我们就完成了钩子，现在可以开始在各个地方使用钩子了。

例如

  const onDataReceived = async (data: BodyInit) => {
    const rawResponse = await fetch('https://backend-endpoint', {
      method: 'POST',
      body: data
    });
    const response = await rawResponse.json();

    setText(response)
  };

  const { isRecording, startRecording, error } = useAudioInput({
    audio: true,
    timeInMillisToStopRecording: 2000,
    timeSlice: 400,
    onDataReceived
  })

第二部分是连接一个节点服务器，它可以与谷歌语音到文本 API 进行通信，我已经附上了我在创建节点方面时引用的文档。
https://codelabs.developers.google.com/codelabs/cloud-speech-text-node。

// demo node server which connects with google speech to text api endpoint

const express = require('express');
const cors = require('cors');

const speech = require('@google-cloud/speech');

const client = new speech.SpeechClient();

async function convert(audioBlob) {
  const request = {
    config: {
      encoding: 'WEBM_OPUS', // Ensure this matches the format of the audio being sent
      sampleRateHertz: 48000, // This should match the sample rate of your recording
      languageCode: 'en-US'
    },
    audio: {
      content: audioBlob
    }
  };

  const [response] = await client.recognize(request);

  const transcription = response.results
    .map(result => result.alternatives[0].transcript)
    .join('\n');
  return transcription;
}

const app = express();

app.use(cors())
app.use(express.json());

app.post('/upload', express.raw({ type: '*/*' }), async (req, res) => {
    const audioBlob = req.body;

    const response = await convert(audioBlob);

    res.json(response);
});

app.listen(4000,'0.0.0.0', () => {
  console.log('Example app listening on port 4000!');
});

在本文中，我介绍了如何将音频内容或 blob 发送到 google 语音转文本端点，我们还可以发送 blob uri 而不是内容，唯一的变化是有效负载

// sending url as part of audio object to speech to text api 
...
audio: {url: audioUrl} or audio: {content: audioBlob}
...

与本文相关的代码存在于 Github 中。

以上是通过 Google Speech to Text 进行音频转文本输入的详细内容。更多信息请关注PHP中文网其他相关文章！