亚马逊云AI大语言模型应用下的创新Amazon Transcribe的使用

Transcribe简介

语音识别技术，也被称为自动语音识别（Automatic Speech Recognition，简称ASR），其目标是将人类的语音中的词汇内容转换为计算机可读的输入，例如按键、二进制编码或者字符序列。语音识别技术已经发展了几十年，直到2009年，Hinton把人工智能深度学习解决方案引入语音识别中，语音识别才取得了巨大突破。

Amazon Transcribe 是一项自动语音识别 (ASR) 服务，使开发人员能够轻松地为其应用程序添加语音转文本功能。自从在 re:Invent 2017 发布以来，越来越多的用户将语音识别功能添加到其应用程序和设备中。2019年8月，Amazon Transcribe推出对中文普通话的支持。更加另用户兴奋的是，在中国的北京区域（BJS）和宁夏区域（ZHY）也已支持该项服务。

在日常工作学习中，我们往往需要对一个视频文件增加字幕。传统的方法需要将视频中的对话用文字记录下来，通常采用记事本等工具保存文字记录，然后借助一些工具软件为文字加上时间轴，然后进行人工校对，整个过程需要耗费大量的时间和精力。是否有更快捷的方法呢？下面我们分享一个使用Amazon Transcribe为视频自动增加字幕的示例。

AWS Transcribe 服务于利用机器学习来识别语音文件中的声音，然后将其转化为文本。目前支持英语和西班牙文语音。必须将语音文件保存在S3中，输出结果也会被保存在S3中。

• 输入声音文件，支持 flac、mp3、mp4 和 wav 文件格式。长度不能超过2小时。

• 指定语言。

几个特色功能：

• 发音者识别（speaker identification）：Transcribe 能区别一个语音文件中的多个说话者。支持2到10个发音者

• 支持多声道（channel identification）: 如果声音文件中有多声道

• 支持字典（vocabulary）：比如不能识别的单词，特定领域不常用的单词

体系架构

1. 监测到S3存储桶中的文件变化，触发lambda函数；

2. lambda函数调用Transcribe服务，生成视频对应的文本（json格式）；

3. 对文本进行格式转换，生成字幕文件格式（srt）；

4. 上传字幕文件到存储桶。

控制台操作展示

• 登录自己的AWS账号进入到aws控制管理台， 然后搜索 Transcribe 进入到这个管理后台

• 点击 Create job 按钮就可以使用AWS的语音转文本服务，根据提示添加必备的参数设置

api接口

• StartTranscriptionJob：开始一个转换任务

• ListTranscriptionJobs：获取任务列表

• GetTranscriptionJob：获取任务

• CreateVocabulary：创建字典

• DeleteVocabulary：删除字典

• GetVocabulary：获取字典

• ListVocabularies：获取字典列表

• UpdateVocabulary：上传字典

python使用Transcribe演示实例

类型1

import time
import boto3
​
transcribe = boto3.client(('transcribe'))
job_name = "testTranscribeJob100"
job_uri = "https://s3.dualstack.us-east-1.amazonaws.com/*****/hellosammy.mp3"
​
transcribe.start_transcription_job(TranscriptionJobName=job_name, Media={'MediaFileUri': job_uri}, MediaFormat='mp3', LanguageCode='en-US')
​
while True:
    status = transcribe.get_transcription_job(TranscriptionJobName = job_name)
    if status['TranscriptionJob']['TranscriptionJobStatus'] in ['COMPLETED', "FAILED"]:
        break
​
    print("Job not ready yet...")
    time.sleep(5)
​
print(status)

类型2

• 安装 Python 的程序包

pip3 install boto3
pip3 install amazon_transcribe
pip3 install websocket-client

• import 部分

import hashlib
import hmac
import urllib.parse
from datetime import datetime
import time
import ssl
import json
import websocket
import _thread
from amazon_transcribe.eventstream import EventStreamMessageSerializer
from amazon_transcribe.eventstream import EventStreamBuffer
from boto3.session import Session

• 创建URL 函数

def sign(key, msg):
    return hmac.new(key, msg.encode("utf-8"), hashlib.sha256).digest()
 
def getSignatureKey(key, dateStamp, region, serviceName):
    kDate = sign(("AWS4" + key).encode("utf-8"), dateStamp)
    kRegion = sign(kDate, region)
    kService = sign(kRegion, serviceName)
    kSigning = sign(kService, "aws4_request")
    return kSigning
 
def create_pre_signed_url(region, language_code, media_encoding, sample_rate):
    # 获得access key和secret key
    credentials = Session().get_credentials()
    access_key_id = credentials.access_key
    secret_access_key = credentials.secret_key
 
    method = "GET"
    service = "transcribe"
    endpoint = "wss://transcribestreaming." + region + ".amazonaws.com:8443"
    host = "transcribestreaming." + region + ".amazonaws.com:8443"
    algorithm = "AWS4-HMAC-SHA256"
 
    t = datetime.utcnow()
    amz_date =t.strftime('%Y%m%dT%H%M%SZ')
    datestamp =t.strftime('%Y%m%d')
 
    canonical_uri = "/stream-transcription-websocket"
 
    canonical_headers = "host:" + host + "\n"
    signed_headers = "host"
 
    credential_scope = datestamp + "/" + region + "/" + service + "/" + "aws4_request"
 
    canonical_querystring = "X-Amz-Algorithm=" + algorithm
    canonical_querystring += "&X-Amz-Credential=" + urllib.parse.quote_plus(access_key_id + "/" + credential_scope)
    canonical_querystring += "&X-Amz-Date=" + amz_date
    canonical_querystring += "&X-Amz-Expires=300"
    canonical_querystring += "&X-Amz-SignedHeaders=" + signed_headers
    canonical_querystring += "&language-code="+ language_code +"&media-encoding=" + media_encoding +"&sample-rate=" + sample_rate
 
    # Zero length string for connecting
    payload_hash = hashlib.sha256(("").encode('utf-8')).hexdigest()
 
    canonical_request = method + '\n' \
                        + canonical_uri + '\n' \
                        + canonical_querystring + '\n' \
                        + canonical_headers + '\n' \
                        + signed_headers + '\n' \
                        + payload_hash
 
    string_to_sign = algorithm + "\n" \
                     + amz_date + "\n" \
                     + credential_scope + "\n" \
                     + hashlib.sha256(canonical_request.encode("utf-8")).hexdigest()
 
    signing_key = getSignatureKey(secret_access_key, datestamp, region, service)
 
    signature = hmac.new(signing_key, string_to_sign.encode("utf-8"),
                         hashlib.sha256).hexdigest()
 
    canonical_querystring += "&X-Amz-Signature=" + signature
 
    request_url = endpoint + canonical_uri + "?" + canonical_querystring
 
    return request_url

• main 函数

def main():
    url = create_pre_signed_url("us-east-1", "en-US", "pcm", "16000")
    ws = websocket.create_connection(url, sslopt={"cert_reqs": ssl.CERT_NONE})
 
    _thread.start_new_thread(loop_receiving, (ws,))
    print("Receiving...")
    send_data(ws)
 
    while True:
        time.sleep(1)
main()

• loop_receiving 函数

该函数位于 main 函数上方。它将接收 Amazon Transcribe Streaming Service 的返回数据，并且打印出来。

def loop_receiving(ws):
    try:
        while True:
            result = ws.recv()
 
            if result == '':
                continue
 
            eventStreamBuffer = EventStreamBuffer()
 
            eventStreamBuffer.add_data(result)
            eventStreamMessage = eventStreamBuffer.next()
 
            stream_payload = eventStreamMessage.payload
 
            transcript = json.loads(bytes.decode(stream_payload, "UTF-8"))
 
            print("response:",transcript)
 
            results = transcript['Transcript']['Results']
            if len(results)>0:
                for length in range(len(results)):
                    if 'IsPartial' in results[length]:
                        print('IsPartial:', results[length]['IsPartial'])
 
                    if 'Alternatives' in results[length]:
                        alternatives = results[length]['Alternatives']
                        if len(alternatives)>0:
                            for sublength in range(len(alternatives)):
                                if 'Transcript' in alternatives[sublength]:
                                    print('Transcript:', alternatives[sublength]['Transcript'])
 
 
    except Exception as e:
        if 'WebSocketConnectionClosedException' == e.__class__.__name__:
            print("Error: websocket connection is closed")
        else:
            print(f"Exception Name: {e.__class__.__name__}")

• send_data 函数

该函数位于 main 函数上方。它将发送音频数据到 Amazon Transcribe Streaming Service。其中 testFile 变量是测试音频文件地址，测试音频为 pem 格式，英语，采样率为16000。

def send_data(ws):
 
    testFile = "xxx.pem"
 
    bufferSize = 1024*16
 
    stream_headers = {
        ":message-type": "event",
        ":event-type": "AudioEvent",
        ":content-type": "application/octet-stream",
    }
 
    eventstream_serializer = EventStreamMessageSerializer()
 
    with open(testFile, "rb") as source:
        while True:
            audio_chunk = source.read(bufferSize)
            # 将音频数据进行编码
            event_bytes = eventstream_serializer.serialize(stream_headers, audio_chunk)
 
            ws.send(event_bytes, opcode = 0x2) # 0 x 2 send binary
 
            # end with b'' data bytes
            if len(audio_chunk) == 0:
                break

java使用Transcribe演示实例

import com.amazonaws.AmazonServiceException;
import com.amazonaws.SdkClientException;
import com.amazonaws.auth.AWSCredentials;
import com.amazonaws.auth.AWSStaticCredentialsProvider;
import com.amazonaws.auth.BasicAWSCredentials;
import com.amazonaws.auth.profile.ProfileCredentialsProvider;
import com.amazonaws.regions.Regions;
import com.amazonaws.services.s3.AmazonS3;
import com.amazonaws.services.s3.AmazonS3Client;
import com.amazonaws.services.s3.AmazonS3ClientBuilder;
import com.amazonaws.services.s3.model.*;
import com.amazonaws.services.s3.transfer.TransferManager;
import com.amazonaws.services.s3.transfer.Upload;
import com.amazonaws.services.transcribe.AmazonTranscribe;
import com.amazonaws.services.transcribe.AmazonTranscribeClientBuilder;
import com.amazonaws.services.transcribe.model.*;
import org.omg.CosNaming.NamingContextExtPackage.StringNameHelper;
 
import java.io.*;
import java.util.ArrayList;
import java.util.Arrays;
import java.util.List;
 
/**
 * 代码注意： 生成job只保存到S3,所以要保存到本地的思路是先保存到S3,然后下载,
 * 但是期间需要时间,所以不断的查询是否存在在不断的遍历循环比较慢
 *
 * @author DELL
 * @Desc 将本地MP3上传到S3, 然后再转文本,保存json 文件到本地
 * 基本步骤：
 * 1.建立S3客户端连接
 * 2.上传本地音频到S3库中,并返回一个S3地址
 * 3.上传MP3对应的备用单词词库
 * 3.在Amazon Transcribe 创建一个job,将生成的json 文件保存到MP3同级目录
 */
public class Mp3ToJsonUtils {
 
    // 执行文件配置信息
    private static String FILE_TYPE = "mp3";
    // S3配置信息
    private static String AWS_ACCESS_KEY = "自己生成";
    private static String AWS_SECRET_KEY = "自己生成";
    private static final String BUCKET_NAME = "自己生成";
    private static final String JOB_BUCKET_NAME = "自己生成";
    // Aws对象信息
    private static AmazonS3 s3;
    private static TransferManager tx;
    private static AmazonTranscribe amazonTranscribe;
    private static BasicAWSCredentials awsCredentials;
 
    static {
        //1. 建立连接
        try {
            init_with_key();
        } catch (Exception e) {
            e.printStackTrace();
        }
        awsCredentials = new BasicAWSCredentials(AWS_ACCESS_KEY, AWS_SECRET_KEY);
        amazonTranscribe = AmazonTranscribeClientBuilder.standard().withCredentials(new AWSStaticCredentialsProvider(awsCredentials)).withRegion(Regions.US_EAST_2).build();
    }
 
    public static void main(String[] args) throws Exception {
        List<String> list = new ArrayList<>();
        mp3TOJosn("C:\\Users\\DELL\\Desktop\\BK测试数据\\A_Cinderella_Atlas_5.mp3", list);
    }
 
    public static void mp32Josn(String inPath, String savePath, List<String> list) throws Exception {
        String jsonPath = new File(inPath).getParent();
        String name = new File(inPath).getName().replaceAll(" ", "_").replaceAll("-", "_");
        File file = new File(savePath + "\\" + name + ".json");
        //确保上传一次
        if (file.exists()) {
            System.out.println(savePath + "--->已经处理完毕，跳过处理");
            return;
        }
        //2. 上传文件到S3库,得到上传文件对应的S3 URL
        String s3Path = uploadFileToBucket(inPath, BUCKET_NAME);
        String key = new File(s3Path).getName();
        key = key.replaceAll(" ", "_").replaceAll("-", "_");
        //3. 创建Transcription jobs
        createJob(JOB_BUCKET_NAME, FILE_TYPE, key, s3Path);
        //4. 将json文件下载到本地和MP3 同目录
        // 在创建job 生成json 的时候需要一定的时间,先等待一段时间,然后判断存在不存在
        boolean flag = isObjectExit(BUCKET_NAME, key + ".json");
        while (!flag) {
            flag = isObjectExit(BUCKET_NAME, key + ".json");
        }
        amazonS3Downloading(s3, JOB_BUCKET_NAME, key + ".json", savePath + "\\" + key + ".json");
    }
 
    /**
     * 将 MP3 转成 Json文件保存到本地
     *
     * @param inPath
     * @throws Exception
     */
    public static void mp3TOJosn(String inPath, List<String> list) throws Exception {
        String jsonPath = new File(inPath).getParentFile().getParentFile().getParentFile().getAbsolutePath() + "\\json";
        File file1 = new File(jsonPath);
        if (!file1.exists()) {
            file1.mkdirs();
        }
        mp32Josn(inPath, jsonPath, list);
    }
 
    /**
     * 通过在代码中包含 access key id 和 secret access key 连接 aws
     *
     * @throws Exception
     */
    private static void init_with_key() throws Exception {
        AWSCredentials credentials = null;
        credentials = new BasicAWSCredentials(AWS_ACCESS_KEY, AWS_SECRET_KEY);
        s3 = new AmazonS3Client(credentials);
        //Region usWest2 = Region.getRegion(Regions.US_WEST_2);
        //s3.setRegion(usWest2);
        tx = new TransferManager(s3);
    }
 
    /**
     * 上传一个本地文件(对应位置为path)上传到名为bucketName的bucket
     *
     * @param path       需要上传文件的路径
     * @param bucketName S3中存储文件的桶名称
     *                   return 返回对应上传文件的key
     */
    private static String uploadFileToBucket(String path, String bucketName) {
        String keyName = new File(path).getName();
        File fileToUpload = new File(path);
        if (fileToUpload.exists() == false) {
            System.out.println(path + " not exists!");
            return null;
        }
        PutObjectRequest request = new PutObjectRequest(bucketName, fileToUpload.getName(), fileToUpload);
        Upload upload = tx.upload(request);
        while ((int) upload.getProgress().getPercentTransferred() < 100) {
            try {
                Thread.sleep(1000);
            } catch (InterruptedException e) {
                // TODO Auto-generated catch block
                e.printStackTrace();
            }
        }
        System.out.println(path + " MP3上传成功!");
        String s3Path = "s3://" + BUCKET_NAME + "/" + keyName;
        return s3Path;
    }
 
    /**
     * 创建一个 Transcription jobs
     *
     * @param bucketName     S3桶的名称
     * @param fileName       文件类型 eg: mp3, mp4
     * @param jobName        要创建的job的名称
     * @param S3Path         对应S3桶中的对应MP3 或其他路径的S3URL
     * @param vocabularyName Custom vocabulary 对应的名称
     */
    protected static void createJob(String bucketName, String fileName, String jobName, String S3Path) {
        StartTranscriptionJobRequest startTranscriptionJobRequest = new StartTranscriptionJobRequest();
        Media media = new Media();
        media.setMediaFileUri(S3Path);
        // 设置JOb的相应参数  sampling rate 采（抽）样率;
        startTranscriptionJobRequest.withMedia(media)
                .withLanguageCode(LanguageCode.EnUS)
                .withMediaFormat(fileName)
                .withOutputBucketName(bucketName)
                .withSettings(settings)
                .setTranscriptionJobName(jobName);
 
        amazonTranscribe.startTranscriptionJob(startTranscriptionJobRequest);
        GetTranscriptionJobRequest request;
        request = new GetTranscriptionJobRequest();
        request.withTranscriptionJobName(jobName);
        GetTranscriptionJobResult result = amazonTranscribe.getTranscriptionJob(request);
        String status = result.getTranscriptionJob().getTranscriptionJobStatus();
        while (!status.toUpperCase().equals("COMPLETED")) {
            try {
                Thread.sleep(2000);
            } catch (InterruptedException e) {
                e.printStackTrace();
            }
            //System.out.println(status);
            result = amazonTranscribe.getTranscriptionJob(request);
            status = result.getTranscriptionJob().getTranscriptionJobStatus();
            if (status.toUpperCase().equals("FAILED")) {
                System.out.println(result.getTranscriptionJob().getTranscriptionJobName() + "---> is failed");
                System.out.println(result.getTranscriptionJob().getTranscriptionJobName() + "--->" + result.getTranscriptionJob().getFailureReason());
                throw new RuntimeException("transcriobe 失败");
            }
        }
        System.out.println(jobName + "Mp3 Job 生成成功");
    }
    /**
     * 将S3上的文件下载到本地
     *
     * @param s3Client s3客户端
     * @param bucketName 桶名称
     * @param key 文件名称
     * @param targetFilePath 本地路径
     */
    public static void amazonS3Downloading(AmazonS3 s3Client, String bucketName, String key, String targetFilePath) {
        S3Object object = s3Client.getObject(new GetObjectRequest(bucketName, key));
        if (object != null) {
            System.out.println("Content-Type: " + object.getObjectMetadata().getContentType());
            InputStream input = null;
            FileOutputStream fileOutputStream = null;
            byte[] data = null;
            try {
                //获取文件流
                input = object.getObjectContent();
                data = new byte[input.available()];
                int len = 0;
                fileOutputStream = new FileOutputStream(targetFilePath);
                while ((len = input.read(data)) != -1) {
                    fileOutputStream.write(data, 0, len);
                }
                System.out.println(targetFilePath + "json 文件下载文件成功");
            } catch (IOException e) {
                e.printStackTrace();
            } finally {
                if (fileOutputStream != null) {
                    try {
                        fileOutputStream.close();
                    } catch (IOException e) {
                        e.printStackTrace();
                    }
                }
                if (input != null) {
                    try {
                        input.close();
                    } catch (IOException e) {
                        e.printStackTrace();
                    }
                }
            }
        }
    }
 
    /**
     * 判断名为bucketName的bucket里面是否有一个名为key的object
     *
     * @param bucketName
     * @param key
     * @return
     */
    private static boolean isObjectExit(String bucketName, String key) {
        int len = key.length();
        ObjectListing objectListing = s3.listObjects(bucketName);
        String s = new String();
        for (S3ObjectSummary objectSummary : objectListing.getObjectSummaries()) {
            s = objectSummary.getKey();
            int slen = s.length();
            if (len == slen) {
                int i;
                for (i = 0; i < len; i++) {
                    if (s.charAt(i) != key.charAt(i)) {
                        break;
                    }
                }
                if (i == len) {
                    return true;
                }
            }
        }
        return false;
    }
}

效果演示