STT Providers Overview

Speech-to-Text (STT) converts the caller's audio into text that the LLM can understand. Choosing the right STT provider significantly impacts accuracy, latency, and cost.

Supported Providers

Quick Comparison

Latency (Time to First Partial, lower is better):
──────────────────────────────────────────────────────────────────

Deepgram Nova-3   ████░░░░░░░░░░░░░░░░░░░░░░░░░░░░  80ms
Google Chirp      ████████░░░░░░░░░░░░░░░░░░░░░░░░  120ms
Azure Neural      ████████░░░░░░░░░░░░░░░░░░░░░░░░  130ms
ElevenLabs        ██████████░░░░░░░░░░░░░░░░░░░░░░  150ms
AssemblyAI        ████████████░░░░░░░░░░░░░░░░░░░░  180ms
OpenAI Whisper    ████████████████████████████████  300ms

                  0ms              150ms            300ms

Choosing the Right Provider

For Low Latency (Recommended)

Deepgram Nova-3

• Fastest time-to-first-partial (~80ms)
• Excellent for real-time voice agents
• Best endpointing (detects speech completion)
• Smart formatting included

{
  "sttProvider": "deepgram",
  "sttModel": "nova-3"
}

For Indic Languages

Google Chirp 2

• Best accuracy for Hindi, Tamil, Telugu, Bengali
• 125+ languages supported
• Chirp 2 is optimized for telephony

{
  "sttProvider": "google",
  "sttModel": "chirp_2"
}

For Regional Languages

ElevenLabs Scribe

• Excellent for Assamese, Odia, Punjabi
• 99 languages with good regional coverage
• Competitive pricing

{
  "sttProvider": "elevenlabs"
}

For Enterprise

Azure Neural

• Enterprise SLAs and compliance
• Custom speech models available
• Global deployment options

{
  "sttProvider": "azure"
}

Streaming vs Batch

All our STT integrations use streaming for real-time voice agents:

Batch STT (not suitable for voice):
────────────────────────────────────
User speaks for 5 seconds → Wait → Get full transcript

Streaming STT (what we use):
────────────────────────────────────
User speaks → Interim results every 100ms → Final transcript
                    │
                    └── LLM can start preparing response

Interim Results

Interim results allow faster response preparation:

// STT emits interim results as user speaks
type TranscriptEvent struct {
    Text      string
    IsFinal   bool
    Stability float32  // 0.0-1.0, higher = more stable
}

// Example stream for "What is my order status?"
// t=100ms: {Text: "What", IsFinal: false, Stability: 0.8}
// t=200ms: {Text: "What is", IsFinal: false, Stability: 0.85}
// t=300ms: {Text: "What is my", IsFinal: false, Stability: 0.9}
// t=500ms: {Text: "What is my order", IsFinal: false, Stability: 0.92}
// t=700ms: {Text: "What is my order status", IsFinal: true, Stability: 1.0}

Endpointing

Endpointing detects when the user has finished speaking:

Deepgram Endpointing Configuration

{
  "sttProvider": "deepgram",
  "sttConfig": {
    "endpointing": 300,
    "utterance_end_ms": 1000,
    "interim_results": true
  }
}

Language Support Matrix

Cost Comparison

Audio Requirements

Audio Conversion

// Telephony sends μ-law, STT may need Linear16
func mulawToLinear16(mulaw []byte) []int16 {
    linear := make([]int16, len(mulaw))
    for i, sample := range mulaw {
        linear[i] = mulawToLinearSample(sample)
    }
    return linear
}

Error Handling

Handle STT failures gracefully:

func handleSTTError(err error) {
    switch {
    case errors.Is(err, ErrAudioTooQuiet):
        // Ask user to speak louder
        tts.Speak("I'm having trouble hearing you. Could you speak a bit louder?")

    case errors.Is(err, ErrConnectionLost):
        // Reconnect automatically
        stt.Reconnect()

    case errors.Is(err, ErrRateLimited):
        // Use fallback provider
        stt.SwitchProvider("fallback")
    }
}

Next Steps

• Deepgram Configuration - Fastest STT for voice
• Google Chirp - Best for Indic languages
• Azure Speech - Enterprise deployment
• ElevenLabs Scribe - Regional languages