Live Meeting Minutes vs Post-Call Summarization APIs: LecSync, AssemblyAI and Deepgram Compared

A meeting or lecture app can show its summary at two moments: while people are still talking, or after the recording ends. That choice decides more of the build than it seems: which API you call, when you pay for it, how many times the same audio gets processed, and what your client has to keep in memory.

This article compares the live minutes built into the LecSync Realtime Speech API (the digest option) with three common alternatives: AssemblyAI's streaming LLM Gateway hook, AssemblyAI's async Summarization, and Deepgram's Summarization. Every competitor fact below was read from the vendor's public pricing page or docs on October 8, 2026. LecSync facts come from the developers page and the live minutes docs.

The short answer

If users need to see minutes build up during the session, LecSync pushes topic sections over the same WebSocket as the captions, each with a time range and the segment IDs it came from, for $0.006 per audio minute with translation and speakers included. AssemblyAI can also run an LLM on each finished turn of a live stream, but you write the prompt and merge the results. If a summary after the call is enough and the audio is in English, Deepgram or AssemblyAI async is cheaper.

What "live minutes" means here

Live minutes are a structured summary that a speech API produces while the stream is still open, so the client can render notes during the session. Post-call summarization runs only after the audio is complete, usually as part of a file transcription job. The first suits an app that shows notes on screen during a class or call. The second suits an app that emails a recap afterwards.

Side-by-side facts (checked October 8, 2026)

Sources: AssemblyAI pricing, AssemblyAI streaming with LLM Gateway, AssemblyAI Summarization, Deepgram pricing, Deepgram Summarization.

How each option behaves in a live session

LecSync: sections that settle as topics change

Pass "digest": true when you open the stream. Once enough new text has accumulated (2,000 characters by default, and the docs say the service can change that), the server pushes a digest.update. The current topic stays pending and keeps gaining points. When the speakers move on, that section becomes confirmed and a new one opens. At the end of the session the server summarizes whatever text is left, even if it is under the threshold.

Two details matter for the client. Each digest.update carries the full set of sections, so you replace your state on every push and never merge. And startMs / endMs are computed from the transcript segments, not written by the model, so tapping a section can seek the recording to the right place. The speaker and timestamp walkthrough shows how those segments line up.

How often an update lands depends on how fast people speak. For conversational English, 2,000 characters is roughly a couple of minutes of speech (our estimate, assuming about 150 words a minute).

AssemblyAI: an LLM call per turn on the stream

AssemblyAI's streaming API accepts an llm_gateway parameter at connect time. Its guide shows a prompt containing a {{turn}} placeholder and a model such as claude-sonnet-4-6; the response arrives as an LLMGatewayResponse message on the same socket. The real-time meeting assistant example summarizes "after each speaker turn".

This is flexible: you choose the model and the prompt, and you can ask for action items, decisions or any JSON you like. The trade-off is that each call sees one turn. If you want topic-level sections that grow and close, your code keeps the running state, decides when a topic has ended, and maps results back to timestamps. The same guide also shows the other pattern, collecting turns and sending the whole transcript to LLM Gateway when the session ends.

Async summarization: one pass after the audio is complete

AssemblyAI's Summarization runs inside a pre-recorded transcription request and returns chapter summaries with start and end times. The docs mark it as open beta and suggest the medium effort setting for long, multilingual or detail-critical audio. Deepgram's summarize=v2 returns one short summary for the whole file. Its docs mark the feature as unavailable on streaming and English only, and inputs under 50 words come back unsummarized.

Neither fits an app that wants notes on screen during a class or call. Both fit a recap email well.

What one hour costs

Assumptions: one hour of audio, speaker labels on. Token counts are our estimate: about 9,000 spoken words, at AssemblyAI's stated average of 1.3 tokens per English word, comes to roughly 12,000 tokens. Prices are public list prices on October 8, 2026.

The LLM tokens are the smallest line. With GPT-5 mini on LLM Gateway ($0.25 per 1M input tokens, $2.00 per 1M output), a per-turn summary across one hour costs a few cents. Deepgram's summary of a 12,000-token transcript costs well under a cent. The larger costs are elsewhere: paying for the audio twice if you stream captions and then re-transcribe for a summary, and the engineering time to turn per-turn outputs into minutes a user can read.

If you need live captions and a post-call summary on Deepgram, that means a streaming session (Nova-3 Monolingual is $0.0048 a minute on the current promotional rate, plus $0.0020 a minute for streaming diarization) and then either a second pre-recorded pass or your own LLM call on the streamed text. For the full multi-vendor bill with translation added, see the flat versus add-on pricing breakdown.

Where the alternatives are the better pick

• A recap after the call is enough. AssemblyAI async at about $0.25 an hour and Deepgram at about $0.26 both cost less than LecSync's $0.36. If nobody looks at notes during the session, use them.
• You need your own prompt or model. LecSync's minutes have a fixed section shape, and the digest option takes no prompt. AssemblyAI's LLM Gateway lets you pick from its catalog of models and return any structure you define.
• English only, at the lowest streaming rate. AssemblyAI Universal-Streaming at $0.15 an hour and Deepgram's promotional Nova-3 rate are both below $0.36 before you add diarization.
• You need an SLA. LecSync's pricing page says availability is best-effort and there is no SLA today.

Where LecSync fits

LecSync is built for apps where the summary is part of the live screen: a lecture app showing notes beside the captions, or a meeting client showing decisions as they happen. One connect call turns on captions, a translation track, speaker labels and minutes, and the bill is accepted audio minutes × $0.006. The AssemblyAI and OpenAI comparison covers the translation side of that same decision.

One limit to plan for: a dropped connection ends the stream, and the next connect starts fresh with segment numbering and offsets reset, so keep the last digest.update you received on the client.

FAQ

Does LecSync charge extra for live minutes?

No. The developers page lists the live structured summary among the features included in $0.006 per audio minute, alongside transcription, translation, speaker separation, glossary, AI translation refinement and 30-day storage. Turning digest on does not change the rate, and turning it off does not lower it.

How often does LecSync send a digest update?

After about 2,000 characters of new transcript text by default; the docs note the service can configure this. Each push contains the complete list of sections, so the client replaces its state. At the end of the session the remaining text is summarized even if it is below the threshold.

Can Deepgram summarize a live stream?

Not with its built-in feature. Deepgram's Summarization docs mark streaming as unavailable and list English as the supported language. For a live summary you would send the streamed transcript to your own LLM, or run a pre-recorded request after the session ends.

Can AssemblyAI summarize while the stream is running?

Yes, through the llm_gateway connection parameter. It runs the model you choose on each completed turn and returns the result on the same WebSocket. You pay the streaming rate, any diarization add-on, and LLM Gateway token prices, and your app combines per-turn outputs into minutes.

Are LecSync minutes linked to the recording?

Yes. Every section carries startMs, endMs and the segmentIds it was built from. The times are computed from transcript segments, not generated by the model, so they line up with the audio when you store the recording with store: true.

Which option is cheapest for a summary after the call?

At public list prices on October 8, 2026, AssemblyAI async with Summarization at low effort comes to about $0.25 an hour, and Deepgram Nova-3 pre-recorded plus summary tokens is about $0.26. LecSync is $0.36 an hour but adds live captions, translation and minutes during the session.

The short version

Choose by when the summary has to appear. During the session, with a translation track and speaker labels on the same socket, LecSync's digest does it for one rate. After the session, in English, Deepgram and AssemblyAI async cost less. To try it, create a key in the LecSync console, open a stream with "digest": true, and read the live minutes docs for the full message format.