Answer Box
AssemblyAI now offers two non-streaming transcription paths: async, which processes files of any length in the background and bills at $0.21/hr, and sync, which returns a transcript for clips up to 2 minutes in a single HTTP response with a median latency of 134ms and costs $0.45/hr. Both use the same Universal-3.5 Pro model, so accuracy is identical — the difference is delivery speed and feature scope.
The Core Difference: Job vs. Answer
Async transcription works by submitting audio to a background job; you poll or use a webhook to retrieve the finished transcript later. Sync transcription, by contrast, returns the complete transcript in the same HTTP response — no job tracking, no callbacks. That one structural difference drives everything else: async is optimized for throughput and deep analysis on files of any length, while sync is built for speed on short clips when a user or agent is waiting.
Speed Benchmarks and Pricing
For async, processing time ranges from seconds to a few minutes depending on file length and server load, but billing is based solely on audio duration at $0.21 per hour on Universal-3.5 Pro. Sync is designed for clips up to 2 minutes (40 MB max) and returns a transcript at roughly 134ms median latency — faster than the clip took to record. It is priced at $0.45 per hour, the same rate as the Realtime API. The 2-minute cap is intentional: sync is a fast path for short audio, not a replacement for batch processing long recordings.
Accuracy and Feature Tradeoffs
A common concern is whether sync sacrifices accuracy for speed. It does not. Both paths run the same flagship model, so word-level accuracy, punctuation, and casing are identical. The tradeoff is scope: sync excludes speaker diarization, PII redaction, and the full suite of Speech Understanding models (summarization, entity detection, topic detection). Those features require the complete audio file and the processing time that async provides. For a two-minute clip destined for a live UI, the words alone are usually sufficient.
When to Choose Each Path
Reach for async when files are long — meetings, podcasts, call recordings — or when you need deep speech understanding like full-file diarization and summarization, and no one is waiting on the result in real time. It also suits high-volume processing where billing on audio duration keeps costs predictable.
Reach for sync when clips are short (up to 2 minutes) — dictation, voice commands, single utterances — and a person or agent needs the transcript immediately. Sync simplifies integration by requiring only one HTTP call instead of a webhook receiver or polling loop. It works best when you control when audio starts and stops, so you don't need a live stream.
The Bigger Shift in Transcription Design
For years, the transcription decision was binary: fast-but-live (streaming) or accurate-but-delayed (async). Sync breaks that binary by offering a recorded-audio workflow with request/response simplicity and sub-second latency. As you architect a product, the question is no longer just "live or recorded" but also "does someone need this back while waiting?" For short clips, sync eliminates the need to choose between speed and simplicity.
Frequently Asked Questions
What is the difference between sync and async transcription?
Sync returns a finished transcript in the same HTTP response as the request; async submits a background job collected later by polling or webhook. Sync is for short clips where speed matters; async is for files of any length where completeness and deep analysis matter. Both use the same model, so accuracy is equal.
How fast is the Sync API?
AssemblyAI's Sync API returns a transcript in roughly 134ms median (p50) for a short clip — faster than the audio took to record. Async returns a complete transcript in seconds to a few minutes.
When should I use sync instead of async?
Use sync for short clips where a person or agent waits on the result — dictation, voice commands, IVR prompts, voicemail, or voice-agent turns. Use async for long files, non-real-time consumption, or when you need full-file diarization, summarization, or entity detection.
Is sync the same as real-time streaming?
No. Streaming transcribes live audio over a persistent WebSocket, returning partial words mid-speech. Sync is for recorded clips you already have: send one complete short clip over HTTP and get the full transcript back at once.
Does sync sacrifice accuracy?
No. Sync runs the same Universal-3.5 Pro model as async, so word accuracy, punctuation, and casing are identical. The tradeoff is feature scope — sync omits diarization, PII redaction, and Speech Understanding models.
How do I get started?
To try the Sync API, send a POST request with a short audio file (up to 2 minutes, 40 MB) to the sync endpoint. For async, submit audio to the batch endpoint and poll or set up a webhook for the result. Full documentation is available in the AssemblyAI API docs.