Text-to-Speech (TTS) & Audio Generation

The ADK supports generating audio output directly from compatible language models, such as gemini-2.5-flash-preview-tts. You can configure your agents to respond with audio, specify voice configurations, and handle the multimodal audio response within the standard ADK event stream.

Enabling Audio Output

To enable audio generation, you must configure the agent's GenerateContentConfig with the appropriate ResponseModalities and SpeechConfig.

Minimal TTS Configuration

using GoogleAdk.Core;
using GoogleAdk.Core.Agents;
using GoogleAdk.Core.Abstractions.Models;

var agent = new LlmAgent(new LlmAgentConfig
{
    Name = "AudioAgent",
    // Use a model that supports TTS
    Model = "gemini-2.5-flash-preview-tts",
    Instruction = "You are an enthusiastic voice assistant.",
    // Recommended: disable automatic identity injection for TTS models
    // to avoid potential API errors depending on the model's support for System Instructions.
    DisableIdentity = true, 
    GenerateContentConfig = new GenerateContentConfig
    {
        // Specify that we want AUDIO output
        ResponseModalities = new List<Modality> { Modality.AUDIO },
        SpeechConfig = new SpeechConfig
        {
            VoiceConfig = new VoiceConfig
            {
                // Specify a prebuilt voice (e.g., "Aoede", "Puck", "Charon", "Kore", "Fenrir", "Leda")
                PrebuiltVoiceConfig = new PrebuiltVoiceConfig { VoiceName = "Aoede" }
            }
        }
    }
});

Important Notes on TTS Models

Some TTS-specific models (like gemini-2.5-flash-preview-tts) have specific API constraints: 1. System Instructions: The API might reject standard ChatRole.System messages. The ADK's MeaiLlm wrapper handles this internally by detecting -tts in the model name and prepending your configured Instruction to the user's first prompt. 2. Agent Identity: It's highly recommended to set DisableIdentity = true in LlmAgentConfig to prevent the ADK from auto-injecting its standard "You are an agent. Your internal name is..." prompt, which can interfere with the audio output quality or trigger API errors on strictly constrained models. 3. Multi-turn Chat: Some preview TTS endpoints may only support single-turn requests.

Handling the Audio Response

When a TTS model generates an audio response, the ADK event stream will yield an Event where the Content.Parts list contains an InlineData part with the mime type audio/pcm or audio/wav.

Here is an example of running the agent and saving the generated audio to a file:

using System.IO;

var content = new Content 
{ 
    Role = "user", 
    Parts = new List<Part> { new Part { Text = "Say cheerfully: Have a wonderful day!" } } 
};

// Use an ephemeral or standard runner
var resultEnumerable = runner.RunEphemeralAsync("user", content);

await foreach (var evt in resultEnumerable)
{
    if (evt.Content?.Parts != null)
    {
        foreach (var part in evt.Content.Parts)
        {
            // Check for InlineData representing audio
            if (part.InlineData != null && part.InlineData.MimeType.StartsWith("audio/"))
            {
                // The audio data is base64 encoded
                var audioBytes = Convert.FromBase64String(part.InlineData.Data);

                // Write the raw PCM or WAV data to a file
                await File.WriteAllBytesAsync("output_audio.wav", audioBytes);
                Console.WriteLine($"Audio successfully saved ({audioBytes.Length} bytes)");
            }
            // Some models may return text alongside the audio
            else if (part.Text != null)
            {
                Console.WriteLine($"Text Response: {part.Text}");
            }
        }
    }
}

Sample Application

A complete, runnable example of the Audio Agent can be found in the GoogleAdk/samples/GoogleAdk.Samples.AudioAgent project within the repository.