DEV Community: Neeraj Sharma

Showing live voice agent transcripts in a Flutter app with LiveKit

Neeraj Sharma — Fri, 02 Oct 2026 13:53:20 +0000

If you build a voice agent with LiveKit Agents and a Flutter frontend, at some point you want the conversation on screen: what the user said and what the agent is saying, updating live. The docs cover it, but the pieces are spread out and the older API in many examples is deprecated. Here's what works with the current SDKs.

Where the text comes from

When AgentSession runs STT, it publishes transcriptions to the room on the lk.transcription text stream topic. The agent's own speech goes out on the same topic, synced with audio playback, so it can show up word by word as the agent talks. If the user interrupts, the agent text is cut to what was actually spoken.

The sender identity on each stream is the participant who was transcribed, so you can tell user and agent apart without extra metadata.

The older TranscriptionReceived event and publish_transcription() are deprecated. They use a separate delivery path, so anything written to lk.transcription never reaches a TranscriptionEvent listener. Build on text streams.

The Flutter handler

import 'dart:convert';

final transcript = <String, ({String speaker, String text, bool isFinal})>{};

void listenForTranscripts(Room room) {
  final me = room.localParticipant?.identity;

  room.registerTextStreamHandler('lk.transcription',
      (TextStreamReader reader, String participantIdentity) async {
    final info = reader.info!;
    if (info.attributes['lk.transcribed_track_id'] == null) return; // chat message, not a transcript

    final segmentId = info.attributes['lk.segment_id'] ?? info.id;
    final speaker = participantIdentity == me ? 'You' : 'Agent';
    var text = '';

    void update(bool isFinal) {
      transcript[segmentId] = (speaker: speaker, text: text, isFinal: isFinal);
      onTranscriptChanged(); // setState, notifyListeners, etc.
    }

    try {
      await for (final chunk in reader) {
        text += utf8.decode(chunk.content);
        update(false);
      }
    } catch (_) {
      return; // stream aborted, e.g. during a reconnect
    }

    // read this after the stream closes: agent streams set it in the trailer
    update(info.attributes['lk.transcription_final'] == 'true');
  });
}

The docs example uses reader.readAll(). That waits for the stream to close, and an agent segment stays open until the agent stops talking, so the whole reply lands at once. Reading chunks gives you the word-by-word version.

Interim vs final

User speech and agent speech arrive differently. For user speech you get a new stream for each interim result, then a final one, all with the same lk.segment_id. Key your map by that id and replace the entry every time, or the same sentence shows up three times.

Agent speech is one stream per segment that grows as the agent talks. It opens with lk.transcription_final set to false and flips it to true in the trailer when it closes, which is why the handler checks it after the loop.

If you only want finished sentences, skip entries where isFinal is false.

Register once

registerTextStreamHandler throws if a handler for that topic is already set. In a widget that can rebuild or reconnect, register once after connecting and call room.unregisterTextStreamHandler('lk.transcription') in dispose.

If the transcript still feels laggy after this, look at end-of-turn detection on the agent side. I wrote about tuning that in turn detection and barge-in for voice agents, and if you are estimating what a production voice agent costs per minute, there's a voice AI cost calculator on our site.

The code was checked against the client-sdk-flutter source and the LiveKit docs. Originally published on Axionry Engineering.

Estimate your AI feature's model bill before you build it

Neeraj Sharma — Fri, 02 Oct 2026 11:42:19 +0000

Most AI MVP budgets I see cover the build and forget the monthly line. That line is the model bill, and at a few thousand users it can pass what the build cost.

One formula gets you a usable estimate.

The formula

monthly cost = requests per month x (input tokens x input price + output tokens x output price)

Requests per month is users times how often they use the feature. Input tokens include everything you send: system prompt, retrieved context, chat history. Resending all of that on every request is why input usually dominates.

A worked example

10,000 monthly users, 40 requests each, 4,000 input and 400 output tokens per request. That's 400,000 requests, 1.6B input tokens and 160M output tokens a month.

Model (list price per 1M tokens)	Input	Output	Monthly
Claude Sonnet 5 ($2 / $10)	$3,200	$1,600	$4,800
Gemini 3.1 Flash-Lite ($0.25 / $1.50)	$400	$240	$640

Same feature, 7.5x apart.

Routing and caching

Routing. Most requests don't need the big model. Send the easy 80% to Flash-Lite and the hard 20% to Sonnet, and the $4,800 bill drops to about $1,470.

Caching. If 3,000 of those 4,000 input tokens are the same system prompt and reference docs every time, cache them. Gemini 3.1 Flash-Lite bills cached input at $0.025 per 1M instead of $0.25, which takes the Flash-Lite bill from $640 to about $370 a month. Storage is $0.50 per 1M tokens per hour, so keeping a 3,000-token cache alive all month adds about $1.

Both are build decisions. Adding them after launch means rewriting the request path.

What I'd do before writing code

Write down users, requests per user and token counts per request, even rough ones.
Run the formula for a big and a small model.
If the big-model number scares you, design routing and caching in from day one.

If you want the build cost and the run cost side by side, our AI product cost estimator does both, including the monthly bill at your user count. I also wrote up how we scope and price AI product development in checkpoints.

Prices are list prices from the Anthropic and Google pricing pages, October 2026.

Our voice agent costs 2.5¢ a minute. Here's the whole bill.

Neeraj Sharma — Fri, 02 Oct 2026 11:38:48 +0000

We run an AI interviewer that talks to candidates over phone calls. It started on a managed voice platform at roughly 10¢ a minute. The same interviews now run on a custom LiveKit pipeline at about 2.5¢ a minute, all in. For reference, Retell's pricing calculator defaults to 11¢ a minute and LiveKit Cloud's own calculator to about 4.8¢.

The stack

Telnyx SIP trunk into LiveKit SIP
Silero VAD plus LiveKit's turn detector (v1-mini) running on CPU
xAI streaming STT, Gemini 2.5 Flash-Lite, xAI TTS
Self-hosted agent workers and Redis for session state

Where the money goes

Per minute of an inbound call, at published list prices, assuming 600 characters of agent speech and about 3,000 input / 175 output tokens a minute:

Line item	Rate	Per minute
xAI TTS	$15 per 1M characters	$0.0090
LiveKit SIP fee (Ship plan)	$0.004 per minute	$0.0040
xAI STT (streaming)	$0.20 per hour	$0.0033
Telnyx inbound local	$0.0032 per minute	$0.0032
Gemini 2.5 Flash-Lite	$0.10 in / $0.40 out per 1M tokens	$0.0004
Turn detection (local CPU)	none	$0.0000
Workers, Redis, logs	self-hosted	~$0.0051
Total		~$0.025

A few things I didn't expect:

TTS is the biggest line. Almost a cent a minute, billed by character. Shorter agent turns save more than any model swap would.

The LLM barely matters for cost. Four hundredths of a cent. Pick it for latency and quality, not price.

Rate cards only get you to 2¢. The last half cent is workers, Redis and logging. No vendor lists that.

When it isn't worth building

On per-minute cost alone, building wins easily. Count engineer time and it gets closer. With about $600 a month of fixed infra and two engineer-days a month of upkeep (at $150k a year, 260 working days), a custom pipeline only beats an 11¢ vendor above roughly 20,600 minutes a month. Below that, stay on a managed platform and spend the time on the product.

The full migration, with the numbers behind the 75% cut, is in our 2.5¢ voice AI case study. If you want to run your own volumes, there's a voice AI cost calculator on our site.

Rates are from the xAI, Google, Telnyx and LiveKit pricing pages. Originally published on Axionry Engineering.

Why your NAT gateway bill keeps growing (and the 20 minute fix)

Neeraj Sharma — Fri, 02 Oct 2026 08:22:14 +0000

If your workloads run in private subnets, a big part of your NAT gateway bill is probably traffic to S3. The fix takes a few minutes and costs nothing.

NAT gateway pricing in us-east-1 is $0.045 per hour plus $0.045 per GB processed. The hourly part is small. The per-GB part grows with every image pull, backup and log upload that leaves a private subnet.

Check if this is you

In Cost Explorer, filter Service to "EC2-Other" and group by Usage Type. Look for NatGateway-Bytes. In us-east-1 it has no prefix; other regions get one, like USW2-NatGateway-Bytes. If it's near the top, keep reading.

To see where the bytes go, turn on VPC Flow Logs for the NAT gateway's network interface and sum bytes by destination. If most destinations fall in the S3 ranges from ip-ranges.amazonaws.com/ip-ranges.json, you found it.

Fix 1: S3 and DynamoDB gateway endpoints (free)

Gateway endpoints for S3 and DynamoDB have no hourly or data charge. Create one and attach your private route tables:

aws ec2 create-vpc-endpoint \
  --vpc-id vpc-0abc123 \
  --vpc-endpoint-type Gateway \
  --service-name com.amazonaws.us-east-1.s3 \
  --route-table-ids rtb-0priv1 rtb-0priv2

It only covers buckets in the same region. Creating it drops open TCP connections to S3, so don't do it mid-backup, and check that your clients reconnect. S3 will also see your private IPs instead of the NAT's public IP, so bucket policies that allow by source IP need updating. ECR pulls get cheaper too, because image layers are served from S3.

Fix 2: interface endpoints, but only above break-even

ECR API, STS, CloudWatch Logs and most other services need interface endpoints instead. These are not free: $0.01 per hour per AZ plus $0.01 per GB. Across two AZs that is $14.60 a month (2 x 730 hours x $0.01) before any traffic.

Every GB you move off NAT saves $0.035 ($0.045 minus $0.01). So one endpoint in two AZs pays for itself at about 420 GB a month. Below that, leave the traffic on NAT.

Fix 3: watch cross-AZ traffic to the NAT

With one NAT gateway for the whole VPC, traffic from the other AZs pays inter-AZ transfer just to reach it: $0.01/GB in each direction, so $0.02 for every GB that crosses. A second NAT gateway costs about $33 a month in hours, so it pays off once roughly 1.6 TB a month crosses from that AZ.

Do fix 1 first. It's free and it's usually the biggest chunk.

The full list of leaks I check (NAT, egress, idle EBS, public IPv4 and more) is in this AWS cost optimization checklist, and there's a free cloud cost calculator if you want to plug in your own numbers.

Prices are us-east-1 list prices.