Copy page

Guides

Configuring Agent Voice

The voice configuration controls how your agent sounds (text-to-speech) and how it hears (speech-to-text). This guide covers every provider, model, and tuning parameter available on the platform.

---

Voice Pipeline

Code
Caller speaks → Transcriber (STT) → Brain (LLM) → Voice (TTS) → Caller hears

The transcriber and voice sections of the agent schema control the two ends of this pipeline.

---

Text-to-Speech (TTS)

ElevenLabs

ElevenLabs offers the widest library of natural-sounding voices with fine-grained control over stability and expressiveness.

Code
curl -X PATCH https://api.neuratel.ai/v1/agents/agent_abc123 \
  -H "Authorization: Bearer nk_live_your_api_key" \
  -H "Content-Type: application/json" \
  -d '{
    "voice": {
      "provider": "elevenlabs",
      "model": "eleven_flash_v2_5",
      "voice_id": "gHu9GtaHOXcSqFTK06ux",
      "stability": 0.5,
      "similarity_boost": 0.8,
      "style": 0.0,
      "speed": 1.0
    }
  }'

Code
client.patch("/v1/agents/agent_abc123", json={
    "voice": {
        "provider": "elevenlabs",
        "model": "eleven_flash_v2_5",
        "voice_id": "gHu9GtaHOXcSqFTK06ux",
        "stability": 0.5,
        "similarity_boost": 0.8,
        "style": 0.0,
        "speed": 1.0,
    }
})

Code
await fetch("https://api.neuratel.ai/v1/agents/agent_abc123", {
  method: "PATCH",
  headers: {
    "Authorization": "Bearer nk_live_your_api_key",
    "Content-Type": "application/json",
  },
  body: JSON.stringify({
    voice: {
      provider: "elevenlabs",
      model: "eleven_flash_v2_5",
      voice_id: "gHu9GtaHOXcSqFTK06ux",
      stability: 0.5,
      similarity_boost: 0.8,
      style: 0.0,
      speed: 1.0,
    },
  }),
});

Parameter	Range	Description
stability	0.0–1.0	Higher = more consistent, lower = more expressive
similarity_boost	0.0–1.0	Higher = closer to original voice sample
style	0.0–1.0	Higher = more stylistic variation (increases latency)
speed	0.25–4.0	Speaking rate multiplier

Model: eleven_flash_v2_5 — avg 71 ms latency

---

Cartesia

Cartesia provides the lowest TTS latency on the platform, making it ideal for fast-paced conversations.

Code
curl -X PATCH https://api.neuratel.ai/v1/agents/agent_abc123 \
  -H "Authorization: Bearer nk_live_your_api_key" \
  -H "Content-Type: application/json" \
  -d '{
    "voice": {
      "provider": "cartesia",
      "model": "sonic-3",
      "voice_id": "your-cartesia-voice-id",
      "language": "en",
      "speed": 1.0
    }
  }'

Code
client.patch("/v1/agents/agent_abc123", json={
    "voice": {
        "provider": "cartesia",
        "model": "sonic-3",
        "voice_id": "your-cartesia-voice-id",
        "language": "en",
        "speed": 1.0,
    }
})

Code
await fetch("https://api.neuratel.ai/v1/agents/agent_abc123", {
  method: "PATCH",
  headers: {
    "Authorization": "Bearer nk_live_your_api_key",
    "Content-Type": "application/json",
  },
  body: JSON.stringify({
    voice: {
      provider: "cartesia",
      model: "sonic-3",
      voice_id: "your-cartesia-voice-id",
      language: "en",
      speed: 1.0,
    },
  }),
});

Parameter	Description
voice_id	Cartesia voice identifier
language	ISO 639-1 language code
speed	Speaking rate multiplier (preset name or float)
emotion	List of emotion tags (e.g. ["Happy", "Excited"]) — Sonic 3 only
volume	Volume adjustment float — Sonic 3 only

Model: sonic-3 — avg 37 ms latency

---

Phantom Speech

Neuratel's own TTS engine. No external provider dependency — all processing happens on Neuratel infrastructure.

Code
curl -X PATCH https://api.neuratel.ai/v1/agents/agent_abc123 \
  -H "Authorization: Bearer nk_live_your_api_key" \
  -H "Content-Type: application/json" \
  -d '{
    "voice": {
      "provider": "phantom",
      "model": "phantom-english-speech-preview",
      "voice": "aria"
    }
  }'

Code
client.patch("/v1/agents/agent_abc123", json={
    "voice": {
        "provider": "phantom",
        "model": "phantom-english-speech-preview",
        "voice": "aria",
    }
})

Code
await fetch("https://api.neuratel.ai/v1/agents/agent_abc123", {
  method: "PATCH",
  headers: {
    "Authorization": "Bearer nk_live_your_api_key",
    "Content-Type": "application/json",
  },
  body: JSON.stringify({
    voice: {
      provider: "phantom",
      model: "phantom-english-speech-preview",
      voice: "aria",
    },
  }),
});

Models: phantom-english-speech-preview, phantom-arabic-speech-preview — avg 63 ms latency

Available voices (10):

English	Arabic
aria (F), bella (F), claire (F)	layla (F), nour (F)
alex (M), david (M), marcus (M)	omar (M), tariq (M)

Note: Phantom Speech uses the voice field (voice name) instead of voice_id.

---

Speech-to-Text (STT)

Deepgram

Code
curl -X PATCH https://api.neuratel.ai/v1/agents/agent_abc123 \
  -H "Authorization: Bearer nk_live_your_api_key" \
  -H "Content-Type: application/json" \
  -d '{
    "transcriber": {
      "provider": "deepgram",
      "model": "nova-3",
      "language": "en-US"
    }
  }'

Code
client.patch("/v1/agents/agent_abc123", json={
    "transcriber": {
        "provider": "deepgram",
        "model": "nova-3",
        "language": "en-US",
    }
})

Code
await fetch("https://api.neuratel.ai/v1/agents/agent_abc123", {
  method: "PATCH",
  headers: {
    "Authorization": "Bearer nk_live_your_api_key",
    "Content-Type": "application/json",
  },
  body: JSON.stringify({
    transcriber: {
      provider: "deepgram",
      model: "nova-3",
      language: "en-US",
    },
  }),
});

Model	Best for
nova-3	General-purpose transcription, highest accuracy
nova-3-medical	Medical terminology, clinical conversations

OpenAI

Code
{
  "transcriber": {
    "provider": "openai",
    "model": "gpt-4o-mini-transcribe",
    "language": "en"
  }
}

---

Voice APIs

List voices for a provider

Code
curl https://api.neuratel.ai/v1/voices/elevenlabs \
  -H "Authorization: Bearer nk_live_your_api_key"

Returns all available voices including name, description, and latency characteristics.

Voice previews are available directly in the Agent Studio — click the play button next to any voice in the selector to hear a sample before committing.

---

Latency comparison

Component	Provider	Model	Avg latency
TTS	Cartesia	sonic-3	37 ms
TTS	Phantom	phantom-english-speech-preview	63 ms
TTS	ElevenLabs	eleven_flash_v2_5	71 ms
LLM	Neuratel	phantom	187 ms

For the fastest end-to-end pipeline, combine Phantom Brain + Cartesia sonic-3 + Deepgram nova-3.

---

Choosing the right voice

By use case

Use case	Recommended setup
High-volume support	Cartesia sonic-3 — lowest latency, clean sound
Premium customer experience	ElevenLabs eleven_flash_v2_5 — widest voice library, most natural
Multi-language (10+ langs)	Phantom Speech — 20 voices across 10 languages
Healthcare	Deepgram nova-3-medical for STT, any TTS provider

Tuning stability vs. expressiveness (ElevenLabs)

Stability	Effect
0.0–0.3	Highly expressive, varies between utterances
0.4–0.6	Balanced — recommended starting point
0.7–1.0	Very consistent, sounds more controlled

For customer support, start at stability: 0.5 and similarity_boost: 0.75. For marketing or sales, try stability: 0.3 for a more dynamic delivery.

---

Troubleshooting

Symptom	Likely cause	Fix
Agent voice sounds robotic	Low stability or wrong voice selection	Increase stability to 0.6, try a different voice
Agent speech is too fast for callers	Speed set above 1.2	Lower speed to 1.0
Agent cuts off caller mid-sentence	VAD mode too aggressive	Switch to sensitive VAD mode in the transcriber
Transcription errors on technical terms	General STT model struggling	Use nova-3-medical for medical terms, or add a glossary to the system prompt
Long pause before agent speaks	TTS latency too high	Switch to Cartesia sonic-3 or Phantom Speech

---

Related

• Agent Studio — visual configuration interface
• Building Your First Agent — end-to-end tutorial
• Tools and Actions — give your agent capabilities