Language Support

Edesy voice agents support 50+ languages with automatic provider selection optimized for each language.

Supported Languages

Tier 1 (Excellent Support)

Tier 2 (Good Support)

Tier 3 (Basic Support)

Configuration

Basic Language Setup

{
  "agent": {
    "name": "Hindi Support Agent",
    "language": "hi-IN",
    "llmProvider": "gemini-2.5",
    "sttProvider": "google",
    "sttModel": "chirp_2",
    "ttsProvider": "azure",
    "ttsVoice": "hi-IN-SwaraNeural"
  }
}

Auto-Detection

Enable language detection for multi-lingual support:

{
  "agent": {
    "name": "Multi-lingual Agent",
    "language": "auto",
    "supportedLanguages": ["en-US", "hi-IN", "es"],
    "sttConfig": {
      "enableLanguageDetection": true
    }
  }
}

Provider Recommendations by Language

English (en-US, en-GB)

{
  "sttProvider": "deepgram",
  "sttModel": "nova-3",
  "ttsProvider": "cartesia",
  "llmProvider": "gemini-2.5"
}

Why: Deepgram Nova-3 has the lowest latency and highest accuracy for English. Cartesia provides natural-sounding voices with fast streaming.

Hindi (hi-IN)

{
  "sttProvider": "google",
  "sttModel": "chirp_2",
  "ttsProvider": "azure",
  "ttsVoice": "hi-IN-SwaraNeural",
  "llmProvider": "gemini-2.5"
}

Why: Google Chirp 2 is specifically optimized for Indic languages. Azure has the best Hindi voices.

Available Hindi Voices (Azure)

Tamil (ta-IN)

{
  "sttProvider": "google",
  "sttModel": "chirp_2",
  "ttsProvider": "azure",
  "ttsVoice": "ta-IN-PallaviNeural",
  "llmProvider": "gemini-2.5"
}

Bengali (bn-IN)

{
  "sttProvider": "google",
  "sttModel": "chirp_2",
  "ttsProvider": "azure",
  "ttsVoice": "bn-IN-TanishaaNeural",
  "llmProvider": "gemini-2.5"
}

Assamese (as-IN)

{
  "sttProvider": "elevenlabs",
  "ttsProvider": "azure",
  "ttsVoice": "as-IN-YashicaNeural",
  "llmProvider": "gemini-2.5"
}

Note: Assamese has limited STT support. ElevenLabs Scribe provides the best accuracy.

Language-Specific Configuration

Script Direction

For RTL languages (Arabic, Hebrew):

{
  "language": "ar",
  "ttsConfig": {
    "scriptDirection": "rtl"
  }
}

Number Formatting

Configure how numbers are spoken:

// Language-specific number formatting
type NumberFormat struct {
    Language        string
    GroupingSep     string  // "," or "."
    DecimalSep      string  // "." or ","
    SpokenFormat    string  // "one two three" vs "hundred twenty three"
}

var formats = map[string]NumberFormat{
    "en-US": {Language: "en-US", GroupingSep: ",", DecimalSep: ".", SpokenFormat: "natural"},
    "hi-IN": {Language: "hi-IN", GroupingSep: ",", DecimalSep: ".", SpokenFormat: "indian"},
    "de":    {Language: "de", GroupingSep: ".", DecimalSep: ",", SpokenFormat: "natural"},
}

Honorifics

For languages with formal address:

{
  "language": "hi-IN",
  "prompt": "Always use 'आप' (respectful you) instead of 'तुम'. Address customers with 'जी' suffix.",
  "llmConfig": {
    "formalityLevel": "formal"
  }
}

Multi-lingual Agents

Code-Switching Support

Handle users who switch languages mid-conversation:

{
  "agent": {
    "name": "Bilingual Support",
    "primaryLanguage": "en-US",
    "secondaryLanguage": "hi-IN",
    "codeSwitchingEnabled": true,
    "prompt": "You can respond in both English and Hindi. Match the language the customer uses. Hinglish (mixed) is acceptable if the customer uses it."
  }
}

Language Detection

type LanguageDetector struct {
    stt STTProvider
}

func (d *LanguageDetector) Detect(audio []byte) (string, float32) {
    result := d.stt.DetectLanguage(audio)
    return result.Language, result.Confidence
}

// In pipeline
func (p *Pipeline) handleLanguageDetection(audio []byte) {
    lang, confidence := p.langDetector.Detect(audio)

    if confidence > 0.8 && lang != p.currentLanguage {
        p.switchLanguage(lang)
    }
}

func (p *Pipeline) switchLanguage(lang string) {
    // Update STT language
    p.stt.SetLanguage(lang)

    // Update TTS voice
    p.tts.SetVoice(getVoiceForLanguage(lang))

    // Update LLM context
    p.llm.SetLanguageContext(lang)
}

Voice Selection by Language

English Voices

Hindi Voices

Spanish Voices

Accuracy Considerations

Word Error Rate (WER) by Language

Improving Accuracy

1. Custom Vocabulary: Add domain-specific terms

{
  "sttConfig": {
    "keywords": ["Edesy:2", "voice agent:2"]
  }
}

2. Prompt Engineering: Help LLM understand context

"If the user's speech is unclear, ask them to repeat.
Common misheard words in Hindi: 'हां' (yes) and 'नहीं' (no)"

3. Post-Processing: Correct common errors

func correctTranscript(text, language string) string {
    corrections := languageCorrections[language]
    for wrong, right := range corrections {
        text = strings.ReplaceAll(text, wrong, right)
    }
    return text
}

Next Steps

• Hindi Configuration - Detailed Hindi setup
• STT Providers - Provider comparison
• TTS Providers - Voice selection