AssemblyAI for a transcription app: file or streaming, diarization, keyterms

Start from when the audio arrives

AssemblyAI exposes this choice as separate APIs. The pre-recorded API transcribes audio and video files with speaker labels and formatting. The realtime API takes streamed audio and returns transcripts as people speak. A sync API returns finished text for a short clip within the same request.

Jiminny shows why the split matters. Its product team wanted streaming transcription so users get help at the moment a conversation happens, plus speaker labels to tell call participants apart. The vendor-published story reports gains such as at least a 51% rise in customer satisfaction; those are the vendor's and customer's figures, not an independent test.

If your product only needs text after the call, the file API is the simpler fit. Aloware shipped automatic call transcription in six weeks, then used the transcripts for call reviews and QA, according to its vendor story.

1234

What changes the bill and the limits

Pre-recorded audio is priced per hour: the pricing page lists $0.21 per hour for Universal-3.5 Pro and $0.15 per hour for Universal-2. Multichannel files are billed per channel, so a one-hour stereo file counts as two hours; each channel is transcribed independently, which the vendor presents as more accurate for multi-speaker recordings.

Streaming has its own rules. Concurrency is described as free and automatically scaling. On the free plan you can open up to 5 new streaming connections per minute. On pay-as-you-go the starting limit is 100 sessions per minute, rising by 10% whenever you use 70% or more of it. Check how streaming time is metered on the pricing page and in the docs before you model cost for long calls.

1

Model and language: two pages, two answers

The pricing page presents Universal-3.5 Pro for 18 languages and Universal-2 for 99 languages at a lower price. The home page describes Universal-3.6 Pro for recorded audio and a realtime counterpart, each with 32 languages and code-switching, and also claims support for 99+ languages overall. These naming and language differences are not reconciled in the sources, so the documentation must settle which model and language list apply to your account.

Language and timing quality decided Kapwing's switch. Its CTO said the previous API lacked accurate word timing and foreign-language translation, and that AssemblyAI ran experiments with them comparing word error rate and timing accuracy. Dovetail also ran a comparison of providers before choosing. Both are vendor-published accounts; the lesson is to run the same kind of comparison on your own recordings.

156

Diarization: who spoke, live or after the fact

Speaker diarization is available for both pre-recorded and realtime models, and the vendor says it trains its diarization models in-house. Speaker Identification can then replace generic labels such as Speaker A with real names or roles. The pricing page lists add-on features in a separate tab, so check there whether diarization changes your per-hour cost.

Two stories show different uses. Dovetail, which ingests interviews and support calls, reports a roughly 10% improvement in speaker labels in its own comparison. Aloware offers speaker-separated transcripts and speaker sentiment inside its voice analytics product. If your recordings put each speaker on a separate channel, per-channel transcription is an alternative worth testing, at the multichannel billing noted above.

146

Specialised terms and what the output feeds

For the pre-recorded API, the pricing page lists keyterms prompting, custom spelling, filler words, language detection and word-level timestamps. The home page adds Medical and Veterinarian modes with domain vocabulary, and a Legal mode marked as coming soon. The sources I read give no limits on keyterm lists, so read the docs for that.

Word timestamps drive real features. Kapwing uses precise word timings to trim video by editing the transcript and to animate subtitles word by word. CallRail goes the other way: it sends calls through a summarization model that returns bullets, headline, gist or paragraph from a single API call, and adds sentiment analysis on top.

157

Guardrails, data handling and the free tier

A Guardrails API detects and redacts personal data and unsafe content in voice data. Aloware lists PII redaction among the features it shipped with AssemblyAI. On retention, the vendor says streaming keeps no audio or transcripts once you opt out of model training. For pre-recorded audio, uploads are deleted within 24 to 48 hours and transcripts after 30 days by default, with a time-to-live as short as one hour. Paid accounts can opt out of training.

AssemblyAI states a SOC 2 Type II audit and offers a data processing agreement. That agreement says primary processing happens in the United States and sets out a process for notice and objection to new subprocessors, which your privacy team should read.

The free plan needs no credit card and covers up to 185 hours of pre-recorded and 333 hours of streaming transcription. As an editorial suggestion, spend it on a sample of your own noisiest recordings, once as files and once as a stream, and compare speaker labels and domain terms before choosing.

148910