# How Amazon Transcribe works How it works Amazon Transcribe uses machine learning models to convert speech to text. In addition to the transcribed text, transcripts contains data about the transcribed content, including confidence scores and timestamps for each word or punctuation mark. To see an output example, refer to the [Data input and output](how-input.md#how-output) section. For a complete list of features that you can apply to your transcription, refer to the [feature summary](feature-matrix.md). Transcription methods can be separated into two main categories: + **Batch transcriptions**: Transcribe media files that have been uploaded into an Amazon S3 bucket. You can use the [AWS CLI](getting-started-cli.md), [AWS Management Console](getting-started-console.md), and various [AWS SDKs](getting-started-sdk.md) for batch transcriptions. + **Streaming transcriptions**: Transcribe media streams in real time. You can use the [AWS Management Console](getting-started-console.md), [HTTP/2](streaming-setting-up.md#streaming-http2), [WebSockets](streaming-setting-up.md#streaming-websocket), and various [AWS SDKs](getting-started-sdk.md) for streaming transcriptions. Note that feature and language support differs for batch and streaming transcriptions. For more information, refer to [Amazon Transcribe features](feature-matrix.md) and [Supported languages](supported-languages.md). **Topics** + [ # Data input and output ](how-input.md) + [ # Transcribing numbers and punctuation ](how-numbers.md) **API operations to get you started** Batch: [https://docs.aws.amazon.com/transcribe/latest/APIReference/API_StartTranscriptionJob.html](https://docs.aws.amazon.com/transcribe/latest/APIReference/API_StartTranscriptionJob.html) Streaming: [https://docs.aws.amazon.com/transcribe/latest/APIReference/API_StartStreamTranscription.html](https://docs.aws.amazon.com/transcribe/latest/APIReference/API_StartStreamTranscription.html), StartStreamTranscriptionWebSocket # Data input and output Amazon Transcribe takes audio data, as a media file in an Amazon S3 bucket or a media stream, and converts it to text data. If you're transcribing media files stored in an Amazon S3 bucket, you're performing **batch transcriptions**. If you're transcribing media streams, you're performing **streaming transcriptions**. These two processes have different rules and requirements. With batch transcriptions, you can use [Job queueing](job-queueing.md) if you don't need to process all of your transcription jobs concurrently. This allows Amazon Transcribe to keep track of your transcription jobs and process them when slots are available. **Note** Amazon Transcribe may temporarily store your content to continuously improve the quality of its analysis models. See the [Amazon Transcribe FAQ](https://aws.amazon.com/transcribe/faqs/) to learn more. To request the deletion of content that may have been stored by Amazon Transcribe, open a case with [Support](https://aws.amazon.com/contact-us/). **Topics** + [ ## Media formats ](#how-input-audio) + [ ## Audio channels ](#how-input-channels) + [ ## Sample rates ](#how-input-sample-rates) + [ ## Output ](#how-output) ## Media formats Supported media types differ between batch transcriptions and streaming transcriptions, though lossless formats are recommended for both. See the following table for details: | | Batch | Streaming | | --- | --- | --- | | Supported formats | [\[See the AWS documentation website for more details\]](http://docs.aws.amazon.com/transcribe/latest/dg/how-input.html) | [\[See the AWS documentation website for more details\]](http://docs.aws.amazon.com/transcribe/latest/dg/how-input.html) | | Recommended formats | [\[See the AWS documentation website for more details\]](http://docs.aws.amazon.com/transcribe/latest/dg/how-input.html) | [\[See the AWS documentation website for more details\]](http://docs.aws.amazon.com/transcribe/latest/dg/how-input.html) | For best results, use a lossless format, such as FLAC or WAV with PCM 16-bit encoding. **Note** Streaming transcriptions are not supported with all languages. Refer to the 'Data input' column in the [supported languages table](supported-languages.md) for details. ## Audio channels Amazon Transcribe supports single-channel and dual-channel media. Media with more than two channels is not currently supported. If your audio contains multiple speakers on one channel and you want to partition and label each speaker in your transcription output, you can use [Speaker partitioning (diarization)](diarization.md). If your audio contains speech on two separate channels, you can use [Channel identification](channel-id.md) to transcribe each channel separately within your transcript. Both of these options produce one transcript file. **Note** If you don't enable [Speaker partitioning](diarization.md) or [Channel identification](channel-id.md), your transcript text is provided as one continuous section. ## Sample rates With batch transcription jobs, you can choose to provide a sample rate, though this parameter is optional. If you include it in your request, make sure the value you provide matches the actual sample rate in your audio. If you provide a sample rate that doesn't match your audio, your job may fail. With streaming transcriptions, you must include a sample rate in your request. As with batch transcription jobs, make sure the value you provide matches the actual sample rate in your audio. Sample rates for low fidelity audio, such as telephone recordings, typically use 8,000 Hz. For high fidelity audio, Amazon Transcribe supports values between 16,000 Hz and 48,000 Hz. ## Output Transcription output is in JSON format. The first part of your transcript contains the transcript itself in paragraph form, followed by additional data for every word and punctuation mark. The data provided depends on the features you include in your request. At a minimum, your transcript contains the start time, end time, and confidence score for every word. The [following section](#how-it-works-output) shows example output from a basic transcription request that didn't include any additional options or features. All **batch transcripts** are stored in Amazon S3 buckets. You can choose to save your transcript in your own Amazon S3 bucket, or have Amazon Transcribe use a secure default bucket. To learn more about creating and using Amazon S3 buckets, see [Working with buckets](https://docs.aws.amazon.com/AmazonS3/latest/userguide/creating-buckets-s3.html). If you want your transcript stored in an Amazon S3 bucket that you own, specify the bucket's URI in your transcription request. Make sure you give Amazon Transcribe write permissions for this bucket before starting your batch transcription job. If you specify your own bucket, your transcript remains in that bucket until you remove it. If you don't specify an Amazon S3 bucket, Amazon Transcribe uses a secure service-managed bucket and provides you with a temporary URI you can use to download your transcript. Note that temporary URIs are valid for 15 minutes. If you get an `AccessDenied` error when using the provided URI, make a `GetTranscriptionJob` request to get a new temporary URI for your transcript. If you opt for a default bucket, your transcript is deleted when your job expires (90 days). If you want to keep your transcript past this expiration date, you must download it. **Streaming transcripts** are returned via the same method you're using for your stream. **Tip** If you want to convert your JSON output into a turn-by-turn transcript in Word format, see this [GitHub example (for Python3)](https://github.com/aws-samples/amazon-transcribe-output-word-document). This script works with post-call analytics transcripts and standard batch transcripts with diarization enabled. ### Example output Transcripts provide a complete transcription in paragraph form, followed by a word-for-word breakdown, which provides data for every word and punctuation mark. This includes start time, end time, a confidence score, and a type (`pronunciation` or `punctuation`). The following example is from a simple batch transcription job that didn't include any [additional features](feature-matrix.md). With each additional feature that you apply to your transcription request, you get additional data in your transcript output file. Basic batch transcripts contain two main sections: 1. `transcripts`: Contains the entire transcript in one text block. 1. `items`: Contains information on each word and punctuation mark from the `transcripts` section. 1. `audio_segments`: An audio segment is a specific portion of an audio recording that contains uninterrupted spoken language, with minimal pauses or breaks. This segment captures a natural flow of speech and is captured in `audio_segments` with a start time and end time. The `items` element within an audio segment is a sequence of identifiers that correspond to each item within the segment. Each additional feature you include in your transcription request produces additional information in your transcript. ``` { "jobName": "my-first-transcription-job", "accountId": "111122223333", "results": { "transcripts": [ { "transcript": "Welcome to Amazon Transcribe." } ], "items": [ { "id": 0, "start_time": "0.64", "end_time": "1.09", "alternatives": [ { "confidence": "1.0", "content": "Welcome" } ], "type": "pronunciation" }, { "id": 1, "start_time": "1.09", "end_time": "1.21", "alternatives": [ { "confidence": "1.0", "content": "to" } ], "type": "pronunciation" }, { "id": 2, "start_time": "1.21", "end_time": "1.74", "alternatives": [ { "confidence": "1.0", "content": "Amazon" } ], "type": "pronunciation" }, { "id": 3, "start_time": "1.74", "end_time": "2.56", "alternatives": [ { "confidence": "1.0", "content": "Transcribe" } ], "type": "pronunciation" }, { "id": 4, "alternatives": [ { "confidence": "0.0", "content": "." } ], "type": "punctuation" } ], "audio_segments": [ { "id": 0, "transcript": "Welcome to Amazon Transcribe.", "start_time": "0.64", "end_time": "2.56", "items": [ 0, 1, 2, 3, 4 ] } ] }, "status": "COMPLETED" } ``` # Transcribing numbers and punctuation Transcribing numbers Amazon Transcribe automatically adds punctuation to all supported languages, and capitalizes words appropriately for languages that use case distinction in their writing systems. For most languages, numbers are transcribed into their word forms. However, for languages with support for transcribing numbers, Amazon Transcribe treats numbers differently depending on the context in which they're used. For example, if a speaker says "*Meet me at eight-thirty AM on June first at one-hundred Main Street with three-dollars-and-fifty-cents and one-point-five chocolate bars*," this is transcribed as: + Languages with number transcription support: Meet me at 8:30 a.m. on June 1st at 100 Main Street with \$13.50 and 1.5 chocolate bars + All other languages: Meet me at eight thirty a m on June first at one hundred Main Street with three dollars and fifty cents and one point five chocolate bars To view the languages with support for transcribing numbers, refer to [Supported languages and language-specific features](supported-languages.md).