AWS SDK を使用した Amazon Transcribe カスタム語彙を作成し改良する

次のコードサンプルは、以下の操作方法を示しています。

• Amazon S3 に音声ファイルをアップロードします。

• Amazon Transcribe ジョブを実行してファイルを文字起こしし、結果を取得します。

• カスタム語彙を作成して改良し、文字起こしの精度を向上させます。

• カスタム語彙を使ってジョブを実行し、結果を取得します。

Python

SDK for Python (Boto3)

注記

については、こちらを参照してください GitHub。用例一覧を検索し、AWS コードサンプルリポジトリでの設定と実行の方法を確認してください。

ルイス キャロルによる「ジャバウォッキー」の朗読を収録した音声ファイルを文字起こしします。まず、Amazon Transcribe アクションをラップする関数を作成します。

def start_job(
    job_name,
    media_uri,
    media_format,
    language_code,
    transcribe_client,
    vocabulary_name=None,
):
    """
    Starts a transcription job. This function returns as soon as the job is started.
    To get the current status of the job, call get_transcription_job. The job is
    successfully completed when the job status is 'COMPLETED'.

    :param job_name: The name of the transcription job. This must be unique for
                     your AWS account.
    :param media_uri: The URI where the audio file is stored. This is typically
                      in an Amazon S3 bucket.
    :param media_format: The format of the audio file. For example, mp3 or wav.
    :param language_code: The language code of the audio file.
                          For example, en-US or ja-JP
    :param transcribe_client: The Boto3 Transcribe client.
    :param vocabulary_name: The name of a custom vocabulary to use when transcribing
                            the audio file.
    :return: Data about the job.
    """
    try:
        job_args = {
            "TranscriptionJobName": job_name,
            "Media": {"MediaFileUri": media_uri},
            "MediaFormat": media_format,
            "LanguageCode": language_code,
        }
        if vocabulary_name is not None:
            job_args["Settings"] = {"VocabularyName": vocabulary_name}
        response = transcribe_client.start_transcription_job(**job_args)
        job = response["TranscriptionJob"]
        logger.info("Started transcription job %s.", job_name)
    except ClientError:
        logger.exception("Couldn't start transcription job %s.", job_name)
        raise
    else:
        return job



def get_job(job_name, transcribe_client):
    """
    Gets details about a transcription job.

    :param job_name: The name of the job to retrieve.
    :param transcribe_client: The Boto3 Transcribe client.
    :return: The retrieved transcription job.
    """
    try:
        response = transcribe_client.get_transcription_job(
            TranscriptionJobName=job_name
        )
        job = response["TranscriptionJob"]
        logger.info("Got job %s.", job["TranscriptionJobName"])
    except ClientError:
        logger.exception("Couldn't get job %s.", job_name)
        raise
    else:
        return job



def delete_job(job_name, transcribe_client):
    """
    Deletes a transcription job. This also deletes the transcript associated with
    the job.

    :param job_name: The name of the job to delete.
    :param transcribe_client: The Boto3 Transcribe client.
    """
    try:
        transcribe_client.delete_transcription_job(TranscriptionJobName=job_name)
        logger.info("Deleted job %s.", job_name)
    except ClientError:
        logger.exception("Couldn't delete job %s.", job_name)
        raise



def create_vocabulary(
    vocabulary_name, language_code, transcribe_client, phrases=None, table_uri=None
):
    """
    Creates a custom vocabulary that can be used to improve the accuracy of
    transcription jobs. This function returns as soon as the vocabulary processing
    is started. Call get_vocabulary to get the current status of the vocabulary.
    The vocabulary is ready to use when its status is 'READY'.

    :param vocabulary_name: The name of the custom vocabulary.
    :param language_code: The language code of the vocabulary.
                          For example, en-US or nl-NL.
    :param transcribe_client: The Boto3 Transcribe client.
    :param phrases: A list of comma-separated phrases to include in the vocabulary.
    :param table_uri: A table of phrases and pronunciation hints to include in the
                      vocabulary.
    :return: Information about the newly created vocabulary.
    """
    try:
        vocab_args = {"VocabularyName": vocabulary_name, "LanguageCode": language_code}
        if phrases is not None:
            vocab_args["Phrases"] = phrases
        elif table_uri is not None:
            vocab_args["VocabularyFileUri"] = table_uri
        response = transcribe_client.create_vocabulary(**vocab_args)
        logger.info("Created custom vocabulary %s.", response["VocabularyName"])
    except ClientError:
        logger.exception("Couldn't create custom vocabulary %s.", vocabulary_name)
        raise
    else:
        return response



def get_vocabulary(vocabulary_name, transcribe_client):
    """
    Gets information about a custom vocabulary.

    :param vocabulary_name: The name of the vocabulary to retrieve.
    :param transcribe_client: The Boto3 Transcribe client.
    :return: Information about the vocabulary.
    """
    try:
        response = transcribe_client.get_vocabulary(VocabularyName=vocabulary_name)
        logger.info("Got vocabulary %s.", response["VocabularyName"])
    except ClientError:
        logger.exception("Couldn't get vocabulary %s.", vocabulary_name)
        raise
    else:
        return response



def update_vocabulary(
    vocabulary_name, language_code, transcribe_client, phrases=None, table_uri=None
):
    """
    Updates an existing custom vocabulary. The entire vocabulary is replaced with
    the contents of the update.

    :param vocabulary_name: The name of the vocabulary to update.
    :param language_code: The language code of the vocabulary.
    :param transcribe_client: The Boto3 Transcribe client.
    :param phrases: A list of comma-separated phrases to include in the vocabulary.
    :param table_uri: A table of phrases and pronunciation hints to include in the
                      vocabulary.
    """
    try:
        vocab_args = {"VocabularyName": vocabulary_name, "LanguageCode": language_code}
        if phrases is not None:
            vocab_args["Phrases"] = phrases
        elif table_uri is not None:
            vocab_args["VocabularyFileUri"] = table_uri
        response = transcribe_client.update_vocabulary(**vocab_args)
        logger.info("Updated custom vocabulary %s.", response["VocabularyName"])
    except ClientError:
        logger.exception("Couldn't update custom vocabulary %s.", vocabulary_name)
        raise



def list_vocabularies(vocabulary_filter, transcribe_client):
    """
    Lists the custom vocabularies created for this AWS account.

    :param vocabulary_filter: The returned vocabularies must contain this string in
                              their names.
    :param transcribe_client: The Boto3 Transcribe client.
    :return: The list of retrieved vocabularies.
    """
    try:
        response = transcribe_client.list_vocabularies(NameContains=vocabulary_filter)
        vocabs = response["Vocabularies"]
        next_token = response.get("NextToken")
        while next_token is not None:
            response = transcribe_client.list_vocabularies(
                NameContains=vocabulary_filter, NextToken=next_token
            )
            vocabs += response["Vocabularies"]
            next_token = response.get("NextToken")
        logger.info(
            "Got %s vocabularies with filter %s.", len(vocabs), vocabulary_filter
        )
    except ClientError:
        logger.exception(
            "Couldn't list vocabularies with filter %s.", vocabulary_filter
        )
        raise
    else:
        return vocabs



def delete_vocabulary(vocabulary_name, transcribe_client):
    """
    Deletes a custom vocabulary.

    :param vocabulary_name: The name of the vocabulary to delete.
    :param transcribe_client: The Boto3 Transcribe client.
    """
    try:
        transcribe_client.delete_vocabulary(VocabularyName=vocabulary_name)
        logger.info("Deleted vocabulary %s.", vocabulary_name)
    except ClientError:
        logger.exception("Couldn't delete vocabulary %s.", vocabulary_name)
        raise

ラッパー関数を呼び出して、カスタム語彙なしで音声を文字起こししてから、カスタム語彙の異なるバージョンで文字起こしを行うと、よりよい結果が取得できます。

def usage_demo():
    """Shows how to use the Amazon Transcribe service."""
    logging.basicConfig(level=logging.INFO, format="%(levelname)s: %(message)s")

    s3_resource = boto3.resource("s3")
    transcribe_client = boto3.client("transcribe")

    print("-" * 88)
    print("Welcome to the Amazon Transcribe demo!")
    print("-" * 88)

    bucket_name = f"jabber-bucket-{time.time_ns()}"
    print(f"Creating bucket {bucket_name}.")
    bucket = s3_resource.create_bucket(
        Bucket=bucket_name,
        CreateBucketConfiguration={
            "LocationConstraint": transcribe_client.meta.region_name
        },
    )
    media_file_name = ".media/Jabberwocky.mp3"
    media_object_key = "Jabberwocky.mp3"
    print(f"Uploading media file {media_file_name}.")
    bucket.upload_file(media_file_name, media_object_key)
    media_uri = f"s3://{bucket.name}/{media_object_key}"

    job_name_simple = f"Jabber-{time.time_ns()}"
    print(f"Starting transcription job {job_name_simple}.")
    start_job(
        job_name_simple,
        f"s3://{bucket_name}/{media_object_key}",
        "mp3",
        "en-US",
        transcribe_client,
    )
    transcribe_waiter = TranscribeCompleteWaiter(transcribe_client)
    transcribe_waiter.wait(job_name_simple)
    job_simple = get_job(job_name_simple, transcribe_client)
    transcript_simple = requests.get(
        job_simple["Transcript"]["TranscriptFileUri"]
    ).json()
    print(f"Transcript for job {transcript_simple['jobName']}:")
    print(transcript_simple["results"]["transcripts"][0]["transcript"])

    print("-" * 88)
    print(
        "Creating a custom vocabulary that lists the nonsense words to try to "
        "improve the transcription."
    )
    vocabulary_name = f"Jabber-vocabulary-{time.time_ns()}"
    create_vocabulary(
        vocabulary_name,
        "en-US",
        transcribe_client,
        phrases=[
            "brillig",
            "slithy",
            "borogoves",
            "mome",
            "raths",
            "Jub-Jub",
            "frumious",
            "manxome",
            "Tumtum",
            "uffish",
            "whiffling",
            "tulgey",
            "thou",
            "frabjous",
            "callooh",
            "callay",
            "chortled",
        ],
    )
    vocabulary_ready_waiter = VocabularyReadyWaiter(transcribe_client)
    vocabulary_ready_waiter.wait(vocabulary_name)

    job_name_vocabulary_list = f"Jabber-vocabulary-list-{time.time_ns()}"
    print(f"Starting transcription job {job_name_vocabulary_list}.")
    start_job(
        job_name_vocabulary_list,
        media_uri,
        "mp3",
        "en-US",
        transcribe_client,
        vocabulary_name,
    )
    transcribe_waiter.wait(job_name_vocabulary_list)
    job_vocabulary_list = get_job(job_name_vocabulary_list, transcribe_client)
    transcript_vocabulary_list = requests.get(
        job_vocabulary_list["Transcript"]["TranscriptFileUri"]
    ).json()
    print(f"Transcript for job {transcript_vocabulary_list['jobName']}:")
    print(transcript_vocabulary_list["results"]["transcripts"][0]["transcript"])

    print("-" * 88)
    print(
        "Updating the custom vocabulary with table data that provides additional "
        "pronunciation hints."
    )
    table_vocab_file = "jabber-vocabulary-table.txt"
    bucket.upload_file(table_vocab_file, table_vocab_file)
    update_vocabulary(
        vocabulary_name,
        "en-US",
        transcribe_client,
        table_uri=f"s3://{bucket.name}/{table_vocab_file}",
    )
    vocabulary_ready_waiter.wait(vocabulary_name)

    job_name_vocab_table = f"Jabber-vocab-table-{time.time_ns()}"
    print(f"Starting transcription job {job_name_vocab_table}.")
    start_job(
        job_name_vocab_table,
        media_uri,
        "mp3",
        "en-US",
        transcribe_client,
        vocabulary_name=vocabulary_name,
    )
    transcribe_waiter.wait(job_name_vocab_table)
    job_vocab_table = get_job(job_name_vocab_table, transcribe_client)
    transcript_vocab_table = requests.get(
        job_vocab_table["Transcript"]["TranscriptFileUri"]
    ).json()
    print(f"Transcript for job {transcript_vocab_table['jobName']}:")
    print(transcript_vocab_table["results"]["transcripts"][0]["transcript"])

    print("-" * 88)
    print("Getting data for jobs and vocabularies.")
    jabber_jobs = list_jobs("Jabber", transcribe_client)
    print(f"Found {len(jabber_jobs)} jobs:")
    for job_sum in jabber_jobs:
        job = get_job(job_sum["TranscriptionJobName"], transcribe_client)
        print(
            f"\t{job['TranscriptionJobName']}, {job['Media']['MediaFileUri']}, "
            f"{job['Settings'].get('VocabularyName')}"
        )

    jabber_vocabs = list_vocabularies("Jabber", transcribe_client)
    print(f"Found {len(jabber_vocabs)} vocabularies:")
    for vocab_sum in jabber_vocabs:
        vocab = get_vocabulary(vocab_sum["VocabularyName"], transcribe_client)
        vocab_content = requests.get(vocab["DownloadUri"]).text
        print(f"\t{vocab['VocabularyName']} contents:")
        print(vocab_content)

    print("-" * 88)
    print("Deleting demo jobs.")
    for job_name in [job_name_simple, job_name_vocabulary_list, job_name_vocab_table]:
        delete_job(job_name, transcribe_client)
    print("Deleting demo vocabulary.")
    delete_vocabulary(vocabulary_name, transcribe_client)
    print("Deleting demo bucket.")
    bucket.objects.delete()
    bucket.delete()
    print("Thanks for watching!")

• API の詳細については、「AWS SDK for Python (Boto3) API リファレンス」の以下のトピックを参照してください。

  • CreateVocabulary

  • DeleteTranscriptionJob

  • DeleteVocabulary

  • GetTranscriptionJob

  • GetVocabulary

  • ListVocabularies

  • StartTranscriptionJob

  • UpdateVocabulary