Build a Video Call App with Gemini AI Summarization

AI is taking over the world. Resistance is futile. Now, we could either fight the inevitable or succumb to it. In this guide, we will walk through how to combine AI with Agora and build a video call app that uses AI to summarize the call you just had.

Prerequisites

Flutter
A developer account with Agora
Agora Speech-to-Text Server (Can use this example server)
Gemini API Key

Project Setup

Our starting point will be a simple video call app built with Agora. This guide assumes you have a fundamental understanding of how a simple video call works using Agora.

If you do not have a grasp on Agora Fundamentals, you can take a look at the Flutter quickstart guide within the documentation or you could dive deeper with the full Video Call with Agora Flutter course.

This guide will build upon a simple starter video call, which you can find here.

The starter code has a landing screen with only one button that invites you to join a call. This call happens on a single channel called test (it's a demo, okay). You have the remote users' video, your local video, and an end-call button within the call screen. We add and remove the users from the view using the event handlers.

Speech to Text

Agora has a product called Real Time Transcription that you can enable to start transcribing the call of a specific channel.

Real-Time Transcription is a RESTful API that uses an AI microservice to connect to your call and transcribe the spoken audio. This transcription is streamed directly into your video call using the onStreamMessage event. Optionally, it can also be written to a cloud provider, which we will do in this guide as well.

Backend

Real-Time Transcription should be implemented on your business server for a few reasons. With the backend controlling the microservices, you can ensure that only one instance of Real-Time Transcription runs within each channel. You also need to pass your token to the transcription service, so by doing it on the backend, you don't expose that token on the client side.

We will use this server as our backend. This server exposes two endpoints: one for starting the transcription and another for ending it.

Start Real Time Transcription



/start-transcribing/<--Channel Name-->

A successful response will contain the Task ID and the Builder Token, which you must save in your app since you will need to use it to stop the transcription.



{taskId: <--Task ID Value-->, builderToken: <--Builder Token Value-->}

Stop Real Time Transcription



/stop-transcribing/<--Channel Name-->/<--Task ID-->/<--Builder Token-->

Start Transcription within the Call

To make a network call from your Flutter application, you can use the http package. Ensure you use the same App ID on both the app and the backend server. Then, call your API to start the transcribing.

Within the call.dart file, you can add this startTranscription function:



Future<void> startTranscription({required String channelName}) async {
  final response = await post(
    Uri.parse('$serverUrl/start-transcribing/$channelName'),
  );

  if (response.statusCode == 200) {
    print('Transcription Started');
    taskId = jsonDecode(response.body)['taskId'];
    builderToken = jsonDecode(response.body)['builderToken'];
  } else {
    print('Couldn\'t start the transcription : ${response.statusCode}');
  }
}

We will call this function right after our join call method so that it starts as soon as the first user joins the channel. As part of a successful response, you will receive a Task ID and a Builder Token. Save these because you will need to use them to stop the transcription.

When the transcription starts successfully, it acts as a "bot" has joined the call. It's not a real user, but it has its own UID, defined within your backend server. If you are using the server I linked above, the UID is 101. You can exclude this from the remote user's list in the onUserJoined event.



onUserJoined: (RtcConnection connection, int remoteUid, int elapsed) {
  if (remoteUid == 101) return;

  setState(() {
    _remoteUsers.add(remoteUid);
  });
}

End Transcription

To end the transcription, we use a function similar to the starting function. This function will be called stopTranscription and requires us to pass the Task ID and the Builder Token to stop the Real-Time Transcription service.



Future<void> stopTranscription() async {
  final response = await post(
    Uri.parse('$serverUrl/stop-transcribing/$taskId/$builderToken'),
  );
  if (response.statusCode == 200) {
    print('Transcription Stopped');
  } else {
    print('Couldn\'t stop the transcription : ${response.statusCode}');
  }
}

We will call the stopTranscription method in our call screen's dispose method. This will stop the transcription before we leave the channel and release the engine resource.

Retrieve the Transcription

You can access the transcription during the video call by using the onStreamMessage event in the event handler.



onStreamMessage: (RtcConnection connection, int uid, int streamId,
    Uint8List message, int messageType, int messageSize) {
  print(message);
}

You will notice the code above prints out an array of numbers that only mean something to you if you are an all-knowing AI. These numbers are generated using Google's Protocol Buffers (also refered to as protobuf).

Protobufs encode data in a platform-agnostic way. This means that apps or software can retrieve this data and serialize it according to their language.

Decode the Transcription

We will use a Protocol Buffer to decode the message. In this case, we will serialize the random-looking numbers into an object called Message.

Start by creating a .proto file with the following content:



syntax = "proto3";

package call_summary;

message Message {
  int32 vendor = 1;
  int32 version = 2;
  int32 seqnum = 3;
  int32 uid = 4;
  int32 flag = 5;
  int64 time = 6;
  int32 lang = 7;
  int32 starttime = 8;
  int32 offtime = 9;
  repeated Word words = 10;
}
message Word {
  string text = 1;
  int32 start_ms = 2;
  int32 duration_ms = 3;
  bool is_final = 4;
  double confidence = 5;
}

Put this file in a new folder: lib/protobuf/file.proto. This is the input file for the generator to create our Message object.

To use protobuf, you must install the protobuf compiler on your computer. It's available via package managers for Mac (brew install protobuf) and Linux (apt install -y protobuf-compiler). For Windows or if you need a specific version, check the Prottobuf downloads page.

You must also install the protobuf dart package within your project using flutter pub add protobuf.

Now run the following command in your terminal. Four files should be generated in the same lib/protobuf folder.



protoc --proto_path= --dart_out=. lib/protobuf/file.proto

Now that the protobuf is set up, we can use the new Message object to retrieve our transcription in English. This object contains a words array with the transcribed sentences. Using the isFinal variable, we trigger a print statement whenever the sentence finishes.



onStreamMessage: (RtcConnection connection, int uid, int streamId,

    Uint8List message, int messageType, int messageSize) {

  Message text = Message.fromBuffer(message);

  if (text.words[0].isFinal) {

    print(text.words[0].text);

  }

},

Save the Transcription

We have covered the transcription part. Now, we need to get the transcribed text, store it, and then use it to prompt an AI to give us a summary. The Real-Time Transcription service sends the transcribed audio in chunks, so as audio chunks are processed, the serialized data is sent in bursts, each triggering the onStreamMessage event. The simplest way to store it is to concatenate a long string of responses. There are more sophisticated ways to do it, but it is good enough for this demo.

We can hold a string called transcription and add the text as it finalizes.



onStreamMessage: (RtcConnection connection, int uid, int streamId,

    Uint8List message, int messageType, int messageSize) {

  Message text = Message.fromBuffer(message);

  if (text.words[0].isFinal) {

    print(text.words[0].text);

    transcription += text.words[0].text;

  }

},

Get Summary

In the main.dart, we can connect to Gemini using our API key and prompt it to summarize the video call. When you receive this response, you can call setState and update the summary variable to see your changes reflected on the main page.

As I was testing this app, I noticed that the response liked to mention the transcript I passed. Because of this, I added some extra prompts so it does not mention the transcript.

Once we need to pass the transcript string to the retrieveSummary, we'll pass the function to call.dart and call it when the call ends.



late final GenerativeModel model;

@override

void initState() {

  super.initState();

  model = GenerativeModel(model: 'gemini-pro', apiKey: apiKey);

}

void retrieveSummary(String transcription) async {

  final content = [

    Content.text(

      'This is a transcript of a video call that occurred. Please summarize this call in a few sentences. Dont talk about the transcript just give the summary. This is the transcript: $transcription',

    ),

  ];

  final response = await model.generateContent(content);

  setState(() {

    summary = response.text ?? '';

  });

}

Done

With that, we have built an application that triggers a Real-Time Transcription service as soon as someone joins the channel. Then, this transcript is saved on the client side so it can prompt Gemini for a summary and share it with the user.

Congratulations, you are on the path to succumbing to the AI overlords.

You can find the complete code here. And dive into the Real-Time Transcription documentation to build upon this guide.

Thank you for reading!