The Complete Guide to LiveKit Voice AI: STT, TTS, LLM & Realtime Models Compared

Building voice AI applications requires choosing the right combination of models for speech recognition, language understanding, and speech synthesis. LiveKit Agents provides a unified framework for integrating these components, but with dozens of options available, selecting the right stack can be overwhelming.

This guide breaks down every option available in LiveKit, complete with pricing, pros and cons, and recommendations for different use cases.

Table of Contents

• Speech-to-Text (STT) Options
• Text-to-Speech (TTS) Options
• Large Language Model (LLM) Options
• Realtime Model Options
• Recommended Stack for Low Latency + High Accuracy
• Cost Calculator

---

Speech-to-Text (STT) Options

STT models transcribe spoken audio into text in real-time. They’re the first component in the traditional voice AI pipeline.

LiveKit Inference (Managed)

These are available through LiveKit Cloud with automatic billing—no separate API keys needed.

Deepgram

Model	Price/min	Languages
Flux	$0.0077	English only
Nova-3 (mono)	$0.0077	8 languages
Nova-3 (multi)	$0.0092	Multilingual
Nova-2	$0.0058	33 languages
Nova-2 Medical/Phone	$0.0058	English

Pros:

• Excellent accuracy, especially Nova-3
• Built-in phrase endpointing for turn detection (Flux)
• Strong multilingual support
• Medical-specific model available
• Native streaming support

Cons:

• Higher cost than AssemblyAI
• Flux is English-only

AssemblyAI

Model	Price/min	Languages
Universal-Streaming	$0.0025	English
Universal-Streaming-Multilingual	$0.0025	6 languages

Pros:

• Cheapest option at $0.0025/min
• Sophisticated semantic turn detection
• Good accuracy
• Extra params: keyterms_prompt, confidence thresholds

Cons:

• Limited language support (6 languages)
• Fewer model variants than Deepgram

Cartesia (Ink Whisper)

Model	Price/min	Languages
Ink Whisper	$0.0030 ($0.0023 Scale)	98 languages

Pros:

• Widest language support (98 languages)
• Low cost
• Good for international apps

Cons:

• Newer, less proven than Deepgram/AssemblyAI
• No built-in turn detection

ElevenLabs (Scribe V2)

Model	Price/min	Languages
Scribe V2 Realtime	$0.0105	41 languages

Pros:

• Strong multilingual support (41 languages)
• Automatic language detection
• Good for pairing with ElevenLabs TTS

Cons:

• Most expensive option
• No built-in turn detection

Plugin-Based STT Options (Bring Your Own Key)

Provider	Streaming	Turn Detection	Languages	Notes
OpenAI (gpt-4o-transcribe)	No*	No	57+	Very accurate, needs VAD/StreamAdapter
Groq (Whisper)	No*	No	57+	Fast inference, cheap via Groq pricing
Google Cloud	Yes	No	125+	Enterprise, requires GCP account
Azure Speech	Yes	No	100+	Enterprise, good accuracy

*Requires VAD + StreamAdapter for streaming use

STT Recommendation Guide

Use Case	Recommendation
Lowest cost	AssemblyAI ($0.0025/min)
Best accuracy (English)	Deepgram Nova-3 or Flux
Most languages	Cartesia Ink Whisper (98 langs)
Turn detection built-in	Deepgram Flux or AssemblyAI
Enterprise/compliance	Azure or Google Cloud
Budget + multilingual	Cartesia ($0.003/min, 98 langs)

STT Pricing Summary (per minute)

AssemblyAI:     $0.0025  (cheapest)
Cartesia:       $0.0030
Deepgram Nova-2: $0.0058
Deepgram Nova-3: $0.0077
ElevenLabs:     $0.0105  (most expensive)

---

Text-to-Speech (TTS) Options

TTS models convert text into natural-sounding speech. Quality, latency, and voice selection vary significantly between providers.

LiveKit Inference (Managed)

Inworld

Model	Price/1M chars	Languages
Inworld TTS 1	$5.00	12 languages
Inworld TTS 1 Max	$10.00	12 languages

Pros:

• Cheapest option ($5/1M chars)
• Good multilingual support (12 languages)
• Natural, warm voices
• Designed for interactive/gaming use cases

Cons:

• Smaller voice library than ElevenLabs/Cartesia
• No custom voice cloning via Inference
• Less known in voice AI space

Deepgram (Aura)

Model	Price/1M chars	Languages
Aura-1	$15.00	English only
Aura-2	$30.00 ($27 Scale)	English, Spanish

Pros:

• Low cost for English-only apps
• Good for pairing with Deepgram STT (same vendor)
• Professional, natural-sounding voices
• Fast latency

Cons:

• Very limited language support (English + Spanish only)
• Smaller voice selection
• No voice cloning

Rime

Model	Price/1M chars	Languages
Mist V2	$30.00 ($20 Scale)	4 languages
Arcana V2	$40.00 ($30 Scale)	4 languages

Pros:

• Good voice quality
• Distinctive voice personalities (Gen-Z, expressive)
• Good for younger/casual brand voices

Cons:

• Limited language support (en, es, fr, de only)
• Smaller voice library
• Mid-range pricing

Cartesia (Sonic)

Model	Price/1M chars	Languages
Sonic / Sonic-2 / Sonic-3 / Turbo	$50.00 ($37.50 Scale)	15-42 languages

Pros:

• Excellent voice quality
• Strong multilingual support (Sonic-3: 42 languages)
• Emotion controls (excited, sad, etc.)
• Speed and volume adjustments
• Large voice library
• Lowest latency (time-to-first-byte)

Cons:

• Mid-high pricing
• No voice cloning via Inference
• Same price across all model variants

ElevenLabs

Model	Price/1M chars	Languages
Flash v2/v2.5	$150.00 ($60 Scale)	1-32 languages
Turbo v2/v2.5	$150.00 ($60 Scale)	1-32 languages
Multilingual v2	$300.00 ($120 Scale)	29 languages

Pros:

• Best voice quality (most human-like)
• Huge voice library
• Industry-leading expressiveness
• Voice cloning support (via plugin)
• Good multilingual support

Cons:

• Most expensive option by far
• No custom/cloned voices via Inference
• Flash vs Turbo pricing identical

Plugin-Based TTS Options (Bring Your Own Key)

Provider	Voice Cloning	Languages	Notes
OpenAI	No	57+	Simple, good quality, limited voices
Azure Speech	Yes	100+	Enterprise, SSML support
Google Cloud	No	40+	WaveNet/Neural2 voices
Amazon Polly	No	30+	NTTS voices, cost-effective
Hume	No	English	Emotionally expressive AI
LMNT	Yes	English	Fast, low latency
PlayHT	Yes	29+	Good cloning, emotions

TTS Recommendation Guide

Use Case	Recommendation
Lowest cost	Inworld ($5/1M chars)
Best quality	ElevenLabs (premium) or Cartesia
English-only budget	Deepgram Aura-1 ($15/1M)
Most languages	Cartesia Sonic-3 (42 langs)
Voice cloning	ElevenLabs or PlayHT (plugin)
Gaming/interactive	Inworld
Casual/young brand	Rime Arcana
Enterprise/compliance	Azure or Google Cloud
Lowest latency	Cartesia Sonic

TTS Pricing Summary (per 1M characters)

Inworld TTS 1:      $5.00   (cheapest)
Inworld TTS 1 Max:  $10.00
Deepgram Aura-1:    $15.00
Deepgram Aura-2:    $30.00
Rime Mist V2:       $30.00
Rime Arcana:        $40.00
Cartesia Sonic:     $50.00
ElevenLabs Turbo:   $150.00
ElevenLabs Multi:   $300.00 (most expensive)

TTS Cost per Hour of Speech Output

Assuming ~15,000 characters per hour of speech:

Provider	Cost/hour
Inworld TTS 1	~$0.075
Deepgram Aura-1	~$0.23
Cartesia	~$0.75
ElevenLabs Turbo	~$2.25
ElevenLabs Multi	~$4.50

---

Large Language Model (LLM) Options

The LLM is the brain of your voice agent, handling reasoning, responses, and tool orchestration.

LiveKit Inference (Managed)

OpenAI GPT Models

Model	Input/1M	Cached/1M	Output/1M
GPT-4o	$2.50	$1.25	$10.00
GPT-4o mini	$0.15	$0.075	$0.60
GPT-4.1	$2.00	$0.50	$8.00
GPT-4.1 mini	$0.40	$0.10	$1.60
GPT-4.1 nano	$0.10	$0.025	$0.40
GPT-5	$1.25	$0.125	$10.00
GPT-5 mini	$0.25	$0.025	$2.00
GPT-5 nano	$0.05	$0.005	$0.40
GPT-5.1	$1.25	$0.125	$10.00
GPT-5.2	$1.75	$0.175	$14.00

Pros:

• Industry standard, excellent tool calling
• Wide range of model sizes (nano to full)
• Strong reasoning capabilities (GPT-5+)
• Cached input discounts (50-75% off)
• Available via Azure or OpenAI endpoints

Cons:

• Premium pricing for flagship models
• Output tokens more expensive than input

Google Gemini

Model	Input/1M	Cached/1M	Output/1M
Gemini 2.0 Flash Lite	$0.075	N/A	$0.30
Gemini 2.0 Flash	$0.10	N/A	$0.40
Gemini 2.5 Flash Lite	$0.10	$0.01	$0.40
Gemini 2.5 Flash	$0.30	$0.03	$2.50
Gemini 2.5 Pro	$2.50	$0.25	$15.00
Gemini 3 Flash	$0.50	$0.05	$3.00
Gemini 3 Pro	$4.00	$0.40	$18.00

Pros:

• Cheapest options available (2.0 Flash Lite: $0.075 input)
• Excellent multimodal (vision) support
• Strong reasoning at lower cost than GPT
• Good tool calling support
• Huge context windows

Cons:

• Some models still in preview (Gemini 3)
• Less ecosystem tooling than OpenAI

DeepSeek

Model	Input/1M	Output/1M
DeepSeek V3	$0.77	$0.77
DeepSeek V3.2	$0.30	$0.45

Pros:

• Excellent value for quality
• Strong coding abilities
• Good reasoning performance
• Competitive with GPT-4 class models

Cons:

• No cached input pricing
• Fewer model variants
• Limited provider options (Baseten only)

Kimi K2

Model	Input/1M	Output/1M
Kimi K2 Instruct	$0.60	$2.50

Pros:

• Good reasoning capabilities
• Competitive pricing
• Strong at complex tasks

Cons:

• Single model option
• Less established than OpenAI/Google
• No cached input pricing

GPT-OSS 120B (Open Source)

Provider	Input/1M	Cached/1M	Output/1M
Groq	$0.15	$0.075	$0.60
Cerebras	$0.35	N/A	$0.75

Pros:

• Open source model
• Very low cost via Groq
• Fast inference (especially Groq)
• No vendor lock-in

Cons:

• Less capable than proprietary models
• Limited to specific providers

Plugin-Based LLM Options (Bring Your Own Key)

Anthropic Claude

Model	Input/1M	Output/1M	Notes
Claude 3.5 Sonnet	$3.00	$15.00	Best balanced
Claude 3.5 Haiku	$0.80	$4.00	Fast & cheap
Claude 3 Opus	$15.00	$75.00	Most capable

Pros:

• Excellent reasoning and instruction following
• Strong safety features
• Great at nuanced conversations
• Parallel tool calls support

Cons:

• Not available via LiveKit Inference
• Higher output pricing
• Python plugin only

Groq (Direct)

Model	Input/1M	Output/1M
Llama 3.3 70B	$0.59	$0.79
Llama 3.1 8B	$0.05	$0.08
Mixtral 8x7B	$0.24	$0.24

Pros:

• Fastest inference (LPU hardware)
• Open source models
• Very low latency for voice AI
• Great for real-time applications

Cons:

• Limited model selection
• Open source models less capable
• Rate limits on free tier

Other Plugin Options

Provider	Best For
Azure OpenAI	Enterprise, compliance, regional deployment
Amazon Bedrock	AWS ecosystem, Claude access
Mistral AI	European hosting, open weights
Together AI	Wide model selection, fine-tuning
Fireworks	Fast inference, function calling
Ollama	Local/self-hosted, privacy
Perplexity	Search-augmented responses

LLM Recommendation Guide

Use Case	Recommendation
Lowest cost	Gemini 2.0 Flash Lite ($0.075/$0.30)
Best quality	GPT-5.2 or Claude 3.5 Sonnet
Best value	DeepSeek V3.2 or Gemini 2.5 Flash
Fastest inference	Groq (Llama models)
Voice AI optimized	GPT-4.1 mini or Gemini 2.5 Flash
Complex reasoning	GPT-5 or Claude 3 Opus
Budget + good quality	GPT-4o mini or GPT-OSS 120B
Enterprise/compliance	Azure OpenAI
Self-hosted	Ollama

LLM Pricing Summary (per 1M tokens, Input/Output)

BUDGET TIER:
Gemini 2.0 Flash Lite:  $0.075 / $0.30  (cheapest)
GPT-5 nano:             $0.05  / $0.40
GPT-4.1 nano:           $0.10  / $0.40
GPT-OSS 120B (Groq):    $0.15  / $0.60
GPT-4o mini:            $0.15  / $0.60

MID TIER:
DeepSeek V3.2:          $0.30  / $0.45
Gemini 2.5 Flash:       $0.30  / $2.50
GPT-4.1 mini:           $0.40  / $1.60
Kimi K2:                $0.60  / $2.50

PREMIUM TIER:
GPT-5:                  $1.25  / $10.00
GPT-4.1:                $2.00  / $8.00
Gemini 2.5 Pro:         $2.50  / $15.00
GPT-4o:                 $2.50  / $10.00
Claude 3.5 Sonnet:      $3.00  / $15.00
Gemini 3 Pro:           $4.00  / $18.00

LLM Cost per Hour of Voice Conversation

Assuming ~20K input tokens + ~5K output tokens per hour:

Model	Est. Cost/hour
Gemini 2.0 Flash Lite	~$0.003
GPT-5 nano	~$0.003
GPT-OSS 120B (Groq)	~$0.006
GPT-4o mini	~$0.006
DeepSeek V3.2	~$0.008
Gemini 2.5 Flash	~$0.019
GPT-4.1 mini	~$0.016
GPT-4o	~$0.10
Claude 3.5 Sonnet	~$0.12

---

Realtime Model Options

Realtime models process speech directly (speech-to-speech), bypassing the traditional STT→LLM→TTS pipeline. They understand emotional context and verbal cues better than text-based pipelines.

All realtime models are plugin-based (bring your own API key).

OpenAI Realtime API

Model	Audio Input/1M	Audio Output/1M	Text Input/1M	Text Output/1M
gpt-realtime	$32	$64	$5	$20
gpt-realtime-mini	~$10	~$20	$0.60	$2.40
gpt-4o-realtime (legacy)	$100	$200	$5	$20

Per-minute estimate: ~$0.06 input + $0.24 output = ~$0.30/min

Pros:

• Most mature realtime API
• Excellent voice quality and expressiveness
• Multiple voices available (alloy, marin, etc.)
• Semantic VAD for intelligent turn detection
• Text-only mode for use with separate TTS
• Vision support (images/video input)
• Azure OpenAI also available

Cons:

• Most expensive realtime option
• Delayed transcriptions (not real-time)
• History loading can cause text-only responses
• Limited to OpenAI ecosystem

Best for: Premium voice experiences, complex reasoning tasks

Google Gemini Live API

Model	Input (Text)/1M	Output (Text)/1M	Audio
Gemini 2.5 Flash	$0.15	$0.60	~1,500 tokens/min
Gemini 2.0 Flash	$0.10	$0.40	~1,500 tokens/min

Per-minute estimate: ~$0.03-0.05/min (significantly cheaper than OpenAI)

Pros:

• Much cheaper than OpenAI Realtime
• Thinking mode support (Gemini 2.5)
• Affective dialog (emotional responses)
• Proactive audio (model can choose not to respond)
• Multiple voices (Puck, Charon, etc.)
• Text-only mode available
• Vertex AI or Google AI API
• Free tier available during preview

Cons:

• Newer, less mature than OpenAI
• Built-in VAD only (no semantic VAD)
• Fewer voice options than OpenAI
• Preview features may change

Best for: Cost-effective realtime voice, Google Cloud users

Amazon Nova Sonic

Model	Speech Input/1K	Speech Output/1K	Text Input/1K	Text Output/1K
Nova Sonic	$0.0034	$0.0136	$0.00006	$0.00024
Nova 2 Sonic	Similar	Similar	Similar	Similar

Per-minute estimate: ~$0.04-0.06/min (~80% cheaper than OpenAI)

Pros:

• Best price/performance ratio
• Very low latency
• 1M token context window (Nova 2)
• Multiple languages (EN, ES, PT, HI)
• Polyglot voices
• AWS ecosystem integration
• Natural conversation flow

Cons:

• Python only (no Node.js)
• AWS-only (Bedrock)
• Fewer customization options
• VAD-based turn detection only
• Newer platform

Best for: AWS users, cost-sensitive production deployments

xAI Grok Voice Agent API

Pros:

• OpenAI Realtime API compatible
• Built-in X (Twitter) search tool
• Web search capabilities
• File/knowledge base search
• Unique personality options

Cons:

• Python only
• Newer platform
• Limited documentation
• Smaller ecosystem

Best for: Apps needing X/social media integration

Ultravox

Pros:

• All-in-one STT+LLM+TTS
• Simple integration
• Quick setup
• Good for prototyping

Cons:

• Python only
• Less control over individual components
• Smaller community
• Limited customization

Best for: Rapid prototyping, simple voice agents

Realtime Feature Comparison

Feature	OpenAI	Gemini Live	Nova Sonic	xAI Grok	Ultravox
Python	✓	✓	✓	✓	✓
Node.js	✓	✓	—	—	—
Semantic VAD	✓	—	—	✓	—
Text-only mode	✓	✓	—	—	—
Thinking/Reasoning	—	✓	—	—	—
Vision input	✓	✓	—	—	—
Tool calling	✓	✓	✓	✓	✓
Affective dialog	—	✓	—	—	—

Realtime Recommendation Guide

Use Case	Recommendation
Lowest cost	Nova Sonic or Gemini Live
Best quality	OpenAI Realtime
AWS ecosystem	Nova Sonic
Google Cloud	Gemini Live
Social/X integration	xAI Grok
Fastest setup	Ultravox
Production-ready	OpenAI or Nova Sonic
Node.js required	OpenAI or Gemini Live

Realtime Cost Comparison (per 10-minute conversation)

Assuming 5 min user speech + 5 min agent speech:

Provider	Est. Cost
Nova Sonic	~$0.50
Gemini Live	~$0.40-0.50
OpenAI gpt-realtime	~$1.50
OpenAI gpt-4o-realtime	~$3.00

Limitations of All Realtime Models

1. Delayed transcriptions - User transcripts often arrive after agent response
2. No scripted speech - Can’t guarantee exact text output via say()
3. History as text only - Loses emotional context when loading history
4. Turn detection tradeoffs - Built-in VAD may not be as accurate as pipeline STT

Workaround: Use realtime model with modalities=["text"] + separate TTS for full control over speech output while keeping realtime speech understanding.

---

Recommended Stack for Low Latency + High Accuracy

If you’re using Claude (Anthropic) as your LLM—which doesn’t have a realtime speech model—you’ll use the traditional STT → LLM → TTS pipeline. Here’s the optimal configuration:

Best Overall Stack

Component	Recommendation	Why
STT	Deepgram Flux	Fastest STT, semantic turn detection built-in
LLM	Claude 3.5 Sonnet	Best accuracy/speed balance
TTS	Cartesia Sonic-3	Lowest time-to-first-byte
VAD	Silero	Fast, lightweight
Turn Detection	turn_detection="stt"	Uses Flux’s semantic endpointing

from livekit.agents import AgentSession
from livekit.plugins import anthropic, deepgram, cartesia, silero

session = AgentSession(
    stt=deepgram.STT(model="nova-3"),  # or "flux" for English
    llm=anthropic.LLM(
        model="claude-sonnet-4-20250514",
        temperature=0.7,
    ),
    tts=cartesia.TTS(
        model="sonic-english",
        voice="79a125e8-cd45-4c13-8a67-188112f4dd22",  # British Lady
    ),
    vad=silero.VAD.load(),
    turn_detection="stt",  # Use Deepgram's semantic turn detection
)

Why These Choices?

STT: Deepgram Flux/Nova-3

• ~100-200ms to first transcript
• Industry-leading accuracy for English
• Semantic turn detection (knows when user is done thinking, not just pausing)

LLM: Claude 3.5 Sonnet

• ~200-400ms time-to-first-token
• Excellent instruction following, nuanced responses
• Parallel tool calls for complex workflows

TTS: Cartesia Sonic

• ~50-100ms time-to-first-audio (fastest in class)
• Natural, expressive voices
• Emotion controls, speed adjustments

Latency Optimization Tips

1. Use streaming everywhere - All components should stream
2. Sentence-based TTS - Let TTS start on first sentence, not full response
3. Turn detection - turn_detection="stt" reduces false triggers
4. Keep system prompts short - Fewer input tokens = faster LLM response
5. Regional deployment - Deploy agents close to your users

---

Cost Calculator

Recommended Stack Cost (Deepgram Flux + Claude 3.5 Sonnet + Cartesia)

Per-Minute Costs

Component	Rate	Per Minute
Deepgram Flux (STT)	$0.0077/min	$0.0077
Claude 3.5 Sonnet (LLM)	$3 input / $15 output per 1M	~$0.015-0.03
Cartesia Sonic-3 (TTS)	$50/1M chars	~$0.04
Total		~$0.06-0.08/min

Hourly Costs

Scenario	STT	LLM	TTS	Total/Hour
Light conversation (50% talk time)	$0.23	$0.50	$1.20	~$1.93
Active conversation (70% talk time)	$0.32	$0.90	$1.70	~$2.92
Heavy conversation (90% talk time)	$0.42	$1.20	$2.20	~$3.82

Per-Conversation Estimates

Duration	Est. Cost
5 min call	$0.30 - $0.40
10 min call	$0.60 - $0.80
15 min call	$0.90 - $1.20
30 min call	$1.80 - $2.40

Monthly Projections

Usage	Hours/Month	Monthly Cost
Light (1 hr/day)	30 hrs	~$60-90
Medium (4 hrs/day)	120 hrs	~$240-360
Heavy (8 hrs/day)	240 hrs	~$480-720
High volume (24/7)	720 hrs	~$1,440-2,160

Budget Alternative Stack

Component	Budget Choice	Cost
STT	AssemblyAI	$0.0025/min
LLM	Claude 3.5 Haiku	$0.80/$4 per 1M
TTS	Deepgram Aura-2	$30/1M chars

This brings cost down to ~$0.80-1.20/hr with slightly higher latency.

---

Conclusion

Choosing the right voice AI stack depends on your priorities:

• Lowest cost: AssemblyAI + Gemini Flash Lite + Inworld
• Lowest latency: Deepgram Flux + Claude Haiku + Cartesia Sonic
• Best quality: Deepgram Nova-3 + Claude Sonnet + ElevenLabs
• Simplest setup: OpenAI Realtime or Gemini Live (single model)
• AWS ecosystem: Nova Sonic (realtime) or Bedrock models

For most production applications targeting low latency and high accuracy with Claude, the Deepgram + Claude + Cartesia combination offers the best balance at ~$2-3/hour.

---

Last updated: January 2025. Pricing and features may change—always check the official documentation for the latest information.

Resources:

• LiveKit Agents Documentation
• LiveKit Inference Pricing
• OpenAI Pricing
• Anthropic Pricing
• Google Gemini Pricing
• AWS Nova Pricing