How to Transcribe Audio With the AssemblyAI API in Python

November 13, 2026

A working AssemblyAI transcription takes about 20 lines of Python: install the SDK, submit a URL or local file, poll the transcript id until the status reads completed, then read the text field. Free accounts include around $50 in credits, which covers roughly 185 hours at $0.27 per hour, though speaker labels and other add-ons raise the effective rate.

AssemblyAI is an API-first speech company, not a consumer app. There is no upload button. If you want a transcript without writing code, this is not that product.

What does it cost, precisely?

Core transcription is $0.27 per hour of audio on pay-as-you-go. Speaker diarization roughly doubles that, and the audio intelligence features (summarization, sentiment) are priced separately per hour. Volume plans drop the rate; enterprise contracts are negotiated.

Feature Approx. price/hour
Core transcription $0.27
Speaker labels adds ~$0.12
Summarization adds ~$0.12
Auto chapters adds ~$0.12

That $50 free credit sounds generous until you run 40 hours of interviews with diarization and summarization switched on. Do the arithmetic before you batch.

How do you make the first request?

Install the SDK, set the key, submit, poll. The whole sequence is four calls.

  1. pip install assemblyai, then set aai.settings.api_key = "YOUR_KEY"
  2. Build a config with aai.TranscriptionConfig(speaker_labels=True, language_code="en")
  3. Submit with transcriber.transcribe(url, config=config)
  4. Read status, text and utterances from the returned object

The SDK blocks until the transcript finishes, which is convenient for one file and a poor choice for a thousand. For bulk work, submit without waiting and store the transcript id yourself.

Which settings actually change quality?

Three, and only three, matter for most jobs.

  • speaker_labels — required for meetings and interviews. Without it you get a wall of text with no attribution, and you cannot tell who agreed to what.
  • language_code — set it explicitly. Auto-detection works, but it occasionally locks onto the wrong language in the first 30 seconds of a bilingual call and transcribes the rest badly.
  • word_boost — pass your domain terms and product names. Twenty terms is the useful limit; a list of 200 dilutes the effect instead of sharpening it.

Punctuation and formatting are on by default. Turn formatting off if you need raw text for a downstream pipeline that chokes on inserted commas.

How does the polling model work?

Every submission returns a transcript id and a status: queued, processing, completed, or error. Audio length drives turnaround, roughly 0.1-0.3x realtime for typical speech, so a 60-minute file finishes in about 6-20 minutes.

For batch jobs, submit all files first, then poll in a loop — or better, register a webhook and get a POST when each transcript finishes. On a 500-file batch the difference is hours of wasted wall clock.

Rate limits apply per account. Hitting them returns 429s, and the SDK does not retry indefinitely by default, so wrap your calls.

How do you download the final transcript?

The transcript object includes text and, with speaker labels enabled, utterances: a list of objects with speaker, text, start and end in milliseconds.

Export to SRT or VTT using the SDK’s export_subtitles method if you need captions for video. Millisecond timings are good enough for burned-in captions; frame-accurate work still needs manual adjustment at cuts.

Store the raw JSON, not just plain text. Re-processing plain text later loses speaker attribution, and re-running the transcription to recover it costs money twice.

What breaks in production?

Long files. Over about 10 hours, and over roughly 5 GB, submission fails or times out. Split long recordings first.

Poor audio. Heavy background noise, overlapping speakers and phone-quality conference audio all degrade accuracy noticeably. The universal-2 model handles accents far better than earlier versions, but a crosstalk-heavy meeting still produces muddled text.

And URLs. Submitting a remote URL means AssemblyAI fetches it; a signed URL that expires in five minutes can expire while the job sits in the queue. Upload the file to them rather than relying on a third-party host.

How does it compare to Whisper?

OpenAI’s Whisper is free to self-host and costs nothing per hour once you own the GPU. AssemblyAI charges per hour but adds diarization, word-level timestamps and a stability commitment without you running infrastructure.

For occasional transcription, Whisper on a laptop wins. For a product transcribing thousands of hours with reliable speaker labels, the per-hour cost is usually cheaper than the engineering time.

Who is AssemblyAI for?

Developers building transcription into a product: call analytics, meeting notes, podcast pipelines, voice agent evaluation. Teams with volume where a $0.27 per hour marginal cost is easy to model.

Not for individuals wanting one video transcribed, and not for organisations that cannot send audio to a third party.

Quick Summary

  • Core transcription is $0.27 per hour; diarization and summaries add roughly $0.12 each.
  • Free credit is about $50, equal to roughly 185 hours of core transcription.
  • Speaker labels are essential for meetings; set the language explicitly.
  • Use webhooks instead of polling once a batch exceeds a handful of files.
  • Split recordings longer than 10 hours before submitting.