Tutorial 7: Audio Sessions

Build voice-enabled conversations with real-time audio streaming.

What You’ll Learn

• Two modes for voice conversations: VAD and ASM
• Setting up VAD mode with STT and TTS services
• Setting up ASM mode for native audio LLMs
• Handling audio input and output streams
• Turn detection and interruption handling

Prerequisites

• Completed Tutorial 1: First Conversation
• Basic understanding of audio streaming concepts
• API keys for OpenAI or Google (for Gemini Live)

Two Modes for Voice

PromptKit supports two modes for voice conversations:

Mode	Description	Use Case
VAD	Voice Activity Detection pipeline	Standard LLMs (GPT-4, Claude) with separate STT/TTS
ASM	Audio Streaming Model	Native multimodal LLMs (Gemini Live)

VAD Mode: Voice with Text LLMs

VAD mode enables voice conversations with any text-based LLM by adding STT (speech-to-text) and TTS (text-to-speech) stages to the pipeline.

Pipeline Flow

Audio Input → [AudioTurn] → [STT] → [LLM] → [TTS] → Audio Output

Step 1: Create the Pack

Create voice-assistant.pack.json:

{
  "id": "voice-assistant",
  "name": "Voice Assistant",
  "version": "1.0.0",
  "template_engine": {
    "version": "v1",
    "syntax": "{{variable}}"
  },
  "prompts": {
    "assistant": {
      "id": "assistant",
      "name": "Voice Assistant",
      "version": "1.0.0",
      "system_template": "You are a helpful voice assistant. Keep responses concise and natural for spoken conversation. The user's name is {{user_name}}.",
      "parameters": {
        "temperature": 0.7,
        "max_tokens": 150
      }
    }
  }
}

Step 2: Configure VAD Mode

package main

import (
    "context"
    "log"
    "os"

    "github.com/AltairaLabs/PromptKit/sdk"
    "github.com/AltairaLabs/PromptKit/runtime/stt"
    "github.com/AltairaLabs/PromptKit/runtime/tts"
)

func main() {
    // Create STT and TTS services
    sttService := stt.NewOpenAI(os.Getenv("OPENAI_API_KEY"))
    ttsService := tts.NewOpenAI(os.Getenv("OPENAI_API_KEY"))

    // Open duplex conversation with VAD mode
    conv, err := sdk.OpenDuplex("./voice-assistant.pack.json", "assistant",
        sdk.WithVADMode(sttService, ttsService, sdk.DefaultVADModeConfig()),
    )
    if err != nil {
        log.Fatal(err)
    }
    defer conv.Close()

    conv.SetVar("user_name", "Alice")

    // Start audio processing
    ctx := context.Background()
    audioIn, audioOut, err := conv.StartAudio(ctx)
    if err != nil {
        log.Fatal(err)
    }

    // Feed audio from microphone and play output
    // (See complete example for audio I/O implementation)
}

Step 3: Customize VAD Configuration

// Custom VAD settings for different environments
vadConfig := &sdk.VADModeConfig{
    SilenceDuration:   500 * time.Millisecond, // Shorter pause detection
    MinSpeechDuration: 100 * time.Millisecond, // Faster response
    MaxTurnDuration:   20 * time.Second,       // Shorter max turn
    SampleRate:        16000,                  // 16kHz audio
    Language:          "en",                   // English
    Voice:             "nova",                 // TTS voice
    Speed:             1.1,                    // Slightly faster speech
}

conv, _ := sdk.OpenDuplex("./voice-assistant.pack.json", "assistant",
    sdk.WithVADMode(sttService, ttsService, vadConfig),
)

VAD Configuration Options

Option	Default	Description
SilenceDuration	800ms	Silence required to detect turn end
MinSpeechDuration	200ms	Minimum speech before turn can complete
MaxTurnDuration	30s	Maximum turn length
SampleRate	16000	Audio sample rate in Hz
Language	”en”	Language hint for STT
Voice	”alloy”	TTS voice ID
Speed	1.0	TTS speech rate (0.5-2.0)

ASM Mode: Native Audio LLMs

ASM (Audio Streaming Model) mode is for LLMs with native bidirectional audio support, like Gemini Live API. Audio streams directly to and from the model without separate STT/TTS stages.

Pipeline Flow

Audio/Text Input → [DuplexProvider] → Audio/Text Output

Step 1: Configure ASM Mode

package main

import (
    "context"
    "log"
    "os"

    "github.com/AltairaLabs/PromptKit/sdk"
    "github.com/AltairaLabs/PromptKit/runtime/providers"
    "github.com/AltairaLabs/PromptKit/runtime/providers/gemini"
    "github.com/AltairaLabs/PromptKit/runtime/types"
)

func main() {
    ctx := context.Background()

    // Create Gemini streaming session
    session, err := gemini.NewStreamSession(ctx,
        "wss://generativelanguage.googleapis.com/ws/google.ai.generativelanguage.v1alpha.GenerativeService.BidiGenerateContent",
        os.Getenv("GEMINI_API_KEY"),
        &providers.StreamingInputConfig{
            Type:       types.ContentTypeAudio,
            SampleRate: 16000,
            Channels:   1,
        },
    )
    if err != nil {
        log.Fatal(err)
    }

    // Open duplex conversation with ASM mode
    conv, err := sdk.OpenDuplex("./voice-assistant.pack.json", "assistant",
        sdk.WithStreamingConfig(&providers.StreamingInputConfig{
            Type:       types.ContentTypeAudio,
            SampleRate: 16000,
            Channels:   1,
        }),
    )
    if err != nil {
        log.Fatal(err)
    }
    defer conv.Close()

    // Start streaming
    audioIn, audioOut, err := conv.StartAudio(ctx)
    if err != nil {
        log.Fatal(err)
    }

    // Stream audio bidirectionally
    // ...
}

Handling Audio Streams

Sending Audio

// Send audio chunks from microphone
go func() {
    for {
        chunk := readFromMicrophone() // Your audio capture code
        select {
        case audioIn <- chunk:
        case <-ctx.Done():
            return
        }
    }
}()

Receiving Audio

// Play audio output
go func() {
    for chunk := range audioOut {
        if chunk.Error != nil {
            log.Printf("Audio error: %v", chunk.Error)
            continue
        }
        playAudio(chunk.Data) // Your audio playback code
    }
}()

Turn Detection

VAD mode uses turn detection to determine when the user has finished speaking.

Default Behavior

The default turn detector uses silence duration:

• Waits for SilenceDuration of quiet
• Requires at least MinSpeechDuration of speech first
• Forces completion after MaxTurnDuration

Custom Turn Detector

import "github.com/AltairaLabs/PromptKit/runtime/audio"

// Create custom turn detector
detector := audio.NewSilenceDetector(audio.SilenceConfig{
    SilenceThreshold:  500 * time.Millisecond,
    MinSpeechDuration: 100 * time.Millisecond,
})

conv, _ := sdk.OpenDuplex("./voice-assistant.pack.json", "assistant",
    sdk.WithTurnDetector(detector),
    sdk.WithVADMode(sttService, ttsService, nil),
)

Interruption Handling

Users can interrupt the assistant while it’s speaking (barge-in).

// The TTSStageWithInterruption stage handles this automatically
// When speech is detected during TTS output:
// 1. TTS playback stops
// 2. New user speech is processed
// 3. Assistant responds to the interruption

Complete Example: Voice Interview

See the sdk/examples/voice-interview/ directory for a complete working example that includes:

• Audio capture from microphone
• Real-time speech processing
• TTS audio playback
• Turn management
• Interruption handling

Choosing Between VAD and ASM

Consideration	VAD Mode	ASM Mode
LLM Support	Any text LLM	Gemini Live only
Latency	Higher (STT + TTS overhead)	Lower (native audio)
Flexibility	More control over STT/TTS	Less customization
Cost	Separate STT/TTS costs	Single API cost
Quality	Depends on STT/TTS providers	Native quality

Best Practices

1. Sample Rate: Use 16kHz for best compatibility
2. Buffer Size: Keep audio buffers small for low latency
3. Error Handling: Always handle audio errors gracefully
4. Cleanup: Close conversations properly to release resources
5. Testing: Test with various audio inputs and network conditions

What’s Next

• Tutorial 8: TTS Integration - Deep dive into TTS configuration
• Tutorial 9: Variable Providers - Dynamic context injection

See Also

• VADModeConfig Reference - Complete VAD configuration
• TTS Service Reference - TTS providers and options
• Voice Interview Example