Arkinu
AWS service comparison

Amazon Polly vs Amazon Transcribe: differences and when to use each

Amazon Polly
Amazon Transcribe

Often confused on the examQuestions tend to offer one as a distractor for the other: look for the phrase in the scenario that tips the balance.

Machine Learning

Amazon Polly

Service that turns text into lifelike speech (text-to-speech). It is used to build applications that talk, in multiple languages and with different voices, for example for news readers or accessibility.

Exam phrases that point here

  • «convert text into natural-sounding speech»

    Amazon Polly takes text and returns spoken audio with lifelike voices, ready to play or store.

    When not: If the direction is the opposite, from audio to text, the answer is Amazon Transcribe. If you also need to hold a conversation with the user, think of Amazon Lex, which together with Polly forms complete voice interfaces.

  • «control the pronunciation, pauses, or speed of the voice»

    With SSML you mark pauses, emphasis, speed, pitch, and volume, and with custom lexicons you set how specific words are pronounced.

    When not: If the problem is that the system does not recognize certain words correctly in an audio recording, that is speech recognition: it is solved with Amazon Transcribe custom vocabulary.

  • «generate audio files from long texts and store them for reuse»

    Polly asynchronous synthesis processes long texts in the background and places the audio in an Amazon S3 bucket; the generated audio can be played as many times as you want.

    When not: For short sentences that must play almost instantly, use regular synthesis (SynthesizeSpeech), which returns the audio in the response.

  • «highlight each word on screen while the text is read aloud»

    Polly speech marks return the timing of each sentence, word, or viseme (mouth shape), to synchronize highlighted text or animations with the audio.

    When not: Polly does not translate: if the text must be read in another language, it is first translated with Amazon Translate and then synthesized.

  • «read content aloud to improve accessibility»

    Helping people with visual impairments or reading difficulties is one of the typical uses AWS cites for Polly.

    When not: If the content is in an image or a scanned document, the text must first be extracted (Amazon Textract for documents) and then synthesized.

Study Amazon Polly →

Machine Learning

Amazon Transcribe

Automatic speech recognition service that converts audio to text (speech-to-text). It transcribes in real time (streaming) or in batches from files in Amazon S3.

Exam phrases that point here

  • «convert audio or video recordings into text»

    Amazon Transcribe is the AWS automatic speech recognition service: it takes audio and returns the text with timestamps and a confidence score for each word.

    When not: If the direction is the opposite, from text to speech, it is Amazon Polly. If you also need to understand what the user wants and respond in a conversation, it is Amazon Lex.

  • «live captions or real-time transcription»

    Streaming transcription receives the audio in chunks (HTTP/2, WebSocket, or SDK) and returns the text while people speak.

    When not: If the audio is already recorded in S3 and there is no rush, use batch transcription, which supports more formats. Also, not all languages support streaming.

  • «analyze contact center calls: sentiment, issues, and summaries»

    Transcribe Call Analytics is trained for customer service and sales calls: it provides sentiment, call drivers, categories, summaries, and real-time alerts.

    When not: Call Analytics requires two-channel audio (agent and customer separately). If what you already have is text, sentiment is analyzed with Amazon Comprehend.

  • «redact personal data in transcripts»

    Transcribe PII redaction replaces data such as card or Social Security numbers in the transcript, in both batch and streaming.

    When not: Standard redaction does not touch the original audio (Call Analytics can also redact the audio) and does not work for text that does not come from audio: Amazon Comprehend is for that.

  • «recognize technical terms, product names, or industry jargon»

    With a custom vocabulary or a custom language model trained on industry text, Transcribe recognizes domain-specific words better.

    When not: For medical dictation or conversations, there is Amazon Transcribe Medical, which according to the FAQ only supports U.S. English.

Often paired with

  • Amazon S3
Study Amazon Transcribe →