Polly and Transcribe: turn text into speech, then speech back into text

Amazon Polly with Amazon Transcribe, in plain words

Amazon Polly is text-to-speech: you send words, it returns audio. It has four engines with different quality and prices: standard, neural, long-form and generative, the most natural-sounding one. A quick request with SynthesizeSpeech returns audio right away but handles at most 3,000 characters. For longer text, StartSpeechSynthesisTask runs in the background and writes the audio file into an S3 bucket you choose (up to 100,000 billable characters per task).

Amazon Transcribe is the opposite: speech-to-text. A batch job reads an audio or video file from S3 (MP3, WAV, FLAC, MP4 and more), and writes a JSON transcript with the full text plus a timestamp and confidence score for each word. You can name the language, or let Transcribe detect it. If you don't give it an output bucket, the transcript goes to a bucket AWS manages, and you only get a download link that expires after 15 minutes.

Used together through Amazon S3, they make a handy round trip for testing voice features: generate a known sentence with Polly, transcribe it, and compare. In news: on October 20, 2026 AWS upgraded the model behind Polly's Generative engine for more natural speech, added 14 voices (including the US English voices Kenneth and Nate) and brought Hindi and Finnish to that engine for the first time. Existing SSML keeps working, but the docs warn that model updates can slightly change how a voice sounds.

The AI angle

Polly and Transcribe are the ears and mouth around a language model. A typical voice assistant on AWS transcribes what the user says, sends the text to a model in Amazon Bedrock, and has Polly speak the reply. Transcripts are also common inputs for Bedrock summaries or a Knowledge Base, so a meeting recording can become searchable notes. Treat transcripts as untrusted input when you pass them to a model, because the speaker controls what is said.

Hands-on lab: Round trip: Polly speaks a sentence into S3, Transcribe writes it back down

From AWS CloudShell, create a private S3 bucket, have Polly turn one sentence into an MP3 in that bucket with a Generative voice, have Transcribe turn the MP3 back into text in the same bucket, compare the two, and then delete everything.

1. Set names and create a private bucket

   Pick a Region and a unique bucket name. New S3 buckets block public access by default, and nothing in this lab changes that. The sentence is saved to a file so you can compare it later.

   REGION=${AWS_REGION:-us-east-1}
   BUCKET=voice-lab-$(date +%s)-$RANDOM
   JOB=voice-lab-job-$(date +%s)
   echo "Region: $REGION  Bucket: $BUCKET  Job: $JOB"
   
   aws s3 mb s3://$BUCKET --region $REGION
   
   echo 'Welcome to the sandbox. Today we turn this sentence into speech, then back into text, and check that both versions match.' > original.txt

   Done

2. Ask Polly to write the speech into S3

   start-speech-synthesis-task runs in the background and saves the MP3 in your bucket under a key that ends with the task ID. Ruth is one of the US English Generative voices.

   TASK_ID=$(aws polly start-speech-synthesis-task \
     --region $REGION \
     --engine generative \
     --voice-id Ruth \
     --output-format mp3 \
     --output-s3-bucket-name $BUCKET \
     --output-s3-key-prefix speech/ \
     --text file://original.txt \
     --query 'SynthesisTask.TaskId' --output text)
   echo "Polly task: $TASK_ID"

   Done

3. Wait for Polly and find the audio file

   Check the task until it says completed (usually a few seconds; if it says failed, run aws polly get-speech-synthesis-task --region $REGION --task-id $TASK_ID to read TaskStatusReason). The task reports the file's address as OutputUri; the part after the bucket name is the S3 key.

   while true; do
     PSTATUS=$(aws polly get-speech-synthesis-task --region $REGION --task-id $TASK_ID \
       --query 'SynthesisTask.TaskStatus' --output text)
     echo "Polly: $PSTATUS"
     [ "$PSTATUS" = "completed" ] || [ "$PSTATUS" = "failed" ] && break
     sleep 5
   done
   
   URI=$(aws polly get-speech-synthesis-task --region $REGION --task-id $TASK_ID \
     --query 'SynthesisTask.OutputUri' --output text)
   KEY=${URI#*/$BUCKET/}
   echo "Audio key: $KEY"
   aws s3 ls s3://$BUCKET --recursive

   Done

4. Start a Transcribe job on that MP3

   Transcribe reads the file from the same bucket and writes its JSON result under transcripts/. The bucket must be in the same Region as the job, which it is here.

   aws transcribe start-transcription-job \
     --region $REGION \
     --transcription-job-name $JOB \
     --language-code en-US \
     --media MediaFileUri=s3://$BUCKET/$KEY \
     --media-format mp3 \
     --output-bucket-name $BUCKET \
     --output-key transcripts/ \
     --query 'TranscriptionJob.TranscriptionJobStatus' --output text

   Done

5. Wait for Transcribe and read the transcript

   Batch jobs usually finish within a minute for a clip this short. When the status is COMPLETED, download the JSON and print the text next to the original.

   while true; do
     STATUS=$(aws transcribe get-transcription-job --region $REGION --transcription-job-name $JOB \
       --query 'TranscriptionJob.TranscriptionJobStatus' --output text)
     echo "Transcribe: $STATUS"
     [ "$STATUS" = "COMPLETED" ] || [ "$STATUS" = "FAILED" ] && break
     sleep 10
   done
   
   TKEY=$(aws s3api list-objects-v2 --bucket $BUCKET --prefix transcripts/ \
     --query 'Contents[?ends_with(Key, `.json`)].Key | [0]' --output text)
   aws s3 cp s3://$BUCKET/$TKEY transcript.json
   
   echo 'Original:'; cat original.txt
   echo 'Transcript:'
   python3 -c "import json; print(json.load(open('transcript.json'))['results']['transcripts'][0]['transcript'])"

   Done

6. Optional: listen to the voice

   Copy the MP3 into CloudShell, then use Actions > Download file in CloudShell with the path speech.mp3 to play it on your computer.

   aws s3 cp s3://$BUCKET/$KEY speech.mp3
   ls -l speech.mp3

   Done

aws transcribe delete-transcription-job --region $REGION --transcription-job-name $JOB
aws s3 rm s3://$BUCKET --recursive
aws s3 rb s3://$BUCKET
rm -f original.txt transcript.json speech.mp3

# Confirm nothing is left
aws s3 ls | grep "$BUCKET" || echo 'Bucket gone'
aws transcribe list-transcription-jobs --region $REGION --job-name-contains voice-lab \
  --query 'TranscriptionJobSummaries[].TranscriptionJobName'

Gotchas

• Not every voice works with every engine or Region. Asking for a voice the chosen engine doesn't support returns an error, and the Generative engine is offered only in a listed set of Regions. Check the voice list before you hard-code a voice ID.
• Without --output-bucket-name, Transcribe stores the transcript in an AWS-managed bucket for 90 days and hands you a link that expires in 15 minutes. For anything you want to keep, write to your own bucket.
• Transcription job names must be unique in your account and Region, and can't contain spaces. Reusing a name fails with a conflict error, so add a timestamp or delete the old job first.
• Polly's Generative voices don't produce speech marks (the word timing data used for lip-sync or highlighting) and don't have the newscaster style. Use a neural voice if you need those.
• Generative voices can change slightly when AWS updates the model, as it did in October 2026. If you produce audio in batches over months, such as a podcast season, listen for drift between batches.