Voice AI
Voice AI platforms are the APIs, models, and frameworks used to build applications that listen and speak, from speech-to-text and text-to-speech services to complete voice agents that hold real-time conversations on the web or over the phone.
Speech model providers such as Deepgram, AssemblyAI, Speechmatics, Gladia, and Soniox focus on transcription, while ElevenLabs, Cartesia, Rime, Inworld, Fish Audio, and Murf focus on synthesis, and several now offer both. OpenAI, Google, and Amazon serve speech-to-speech models through the OpenAI Realtime API, the Gemini Live API, and Nova Sonic.
Open-source frameworks such as LiveKit Agents and Pipecat connect these services into low-latency pipelines that developers control in code. Hosted platforms such as Vapi, Retell AI, Bland, and Synthflow add telephony, call flows, and analytics for phone agents. Meta acquired PlayAI in 2025, Inworld acquired Ultravox in 2026, and LMNT has shut down.
Voice AI compared
| Platform | Open source | Deployment | Best for |
|---|---|---|---|
| Amazon Nova Sonic | No | API | Speech-to-speech agents on AWS |
| AssemblyAI | No | Managed cloud, API | Transcription plus speech understanding features |
| Azure Speech in Foundry Tools | No | Managed cloud, API | Speech services inside the Azure ecosystem |
| Bland | No | Managed cloud, API, Self-hosted | Phone agents whose calls stay with one vendor |
| Cartesia | No | Managed cloud, API | Low-latency streaming speech for voice agents |
| Deepgram | No | Managed cloud, API | Real-time transcription in production voice apps |
| ElevenLabs | No | Managed cloud, API | Expressive voices and voice agents from one vendor |
| Fish Audio | No | Managed cloud, API, Self-hosted | Voice cloning with an option to self-host models |
| Gemini Live API | No | API | Voice and video agents on Gemini models |
| Gladia | No | Managed cloud, API | Multilingual transcription with audio intelligence add-ons |
| Hume AI | No | Managed cloud, API | Analyzing and evaluating expressive speech |
| Inworld | No | Managed cloud, API | Realtime voice for consumer applications |
| LiveKit Agents | Yes (Apache-2.0) | Library, Managed cloud, Self-hosted | WebRTC voice agents with full code control |
| Murf | No | Managed cloud, API | Multilingual text-to-speech for agents and content |
| OpenAI Realtime API | No | API | Speech-to-speech agents on OpenAI models |
| Pipecat | Yes (BSD-2-Clause) | Library, Managed cloud, Self-hosted | Provider-agnostic voice pipelines in Python |
| Retell AI | No | Managed cloud, API, Self-hosted | Call automation built with a visual flow editor |
| Rime | No | Managed cloud, API | Conversational voices for phone and contact-center agents |
| Soniox | No | Managed cloud, API | Multilingual live transcription and translation |
| Speechmatics | No | Managed cloud, API, Self-hosted | Transcription with on-premises or on-device deployment |
| Synthflow | No | Managed cloud | Contact centers deploying voice agents |
| TEN Framework | No | Library, Self-hosted | Real-time voice agents alongside Agora services |
| Twilio ConversationRelay | No | Managed cloud, API | Adding an LLM to existing Twilio voice numbers |
| Ultravox | No | Managed cloud, API | Speech-native voice agents with telephony |
| Vapi | No | Managed cloud, API | API-driven phone agents with swappable model providers |
Profiles

Amazon Nova Sonic
Amazon Web Services · aws.amazon.com
Amazon Nova 2.5 Sonic is Amazon's speech-to-speech model for building voice agents that hold natural, real-time conversations, with reasoning, instruction following, and tool calling so an agent can take actions during a call.
AssemblyAI
assemblyai.com
AssemblyAI provides speech AI models for speech-to-text on pre-recorded and streaming audio, with speaker detection, sentiment analysis, and other speech understanding features available through its API and SDKs.
Azure Speech in Foundry Tools
Microsoft · azure.microsoft.com
Azure Speech in Foundry Tools, formerly Azure AI Speech, provides speech recognition, text-to-speech, and customized speech models for multilingual applications, and it supplies transcription and synthesis for Voice Live agents in Microsoft Foundry.

Bland
bland.ai
Bland is an enterprise platform for AI phone agents that runs its own speech-to-text, language, and text-to-speech models on its own infrastructure, with on-premises options for sensitive workloads in industries such as healthcare and finance.
Cartesia
cartesia.ai
Cartesia is an AI lab building real-time voice models and agents, offering streaming text-to-speech over WebSockets, real-time speech-to-text with its Ink models, and a platform for production voice agents.
Deepgram
deepgram.com
Deepgram offers speech-to-text, text-to-speech, and Voice Agent APIs for enterprise voice applications, with real-time streaming, SDKs, and a command-line tool for transcribing audio and synthesizing speech.
ElevenLabs
elevenlabs.io · London, United Kingdom · founded 2022
ElevenLabs provides APIs and SDKs for text-to-speech, speech-to-text, voice cloning, sound effects, and voice changing, plus a platform for conversational voice agents, with voices and models that cover many languages.
Fish Audio
fish.audio
Fish Audio offers text-to-speech, voice cloning, and speech-to-text through an API with real-time streaming, and it publishes speech models under the Fish Audio Community License, which permits research and limited commercial use and self-hosting.

Gemini Live API
Google · ai.google.dev
The Gemini Live API enables low-latency, real-time voice and video interactions with Gemini models over a stateful WebSocket connection, with barge-in, audio transcriptions, and tool use including function calling and Google Search.
Gladia
gladia.io
Gladia is an audio transcription and intelligence API for voice products, providing real-time and pre-recorded speech-to-text plus translation, summarization, and entity extraction, with SDKs, a CLI, and an MCP server.
Hume AI
hume.ai
Hume AI works on emotionally intelligent voice AI; its current developer platform centers on an Audio Decisions API that answers plain-language questions about audio and an Expression Measurement API, plus voice evaluation services.

Inworld
inworld.ai
Inworld provides voice APIs for text-to-speech, speech-to-text, and realtime conversations over WebSockets, plus inference APIs for text, image, and video models, and it acquired the Ultravox voice agent platform in 2026.

LiveKit Agents
LiveKit · docs.livekit.io
LiveKit Agents is an open-source framework for realtime voice and multimodal agents that join LiveKit rooms as participants over WebRTC, with Python and Node.js SDKs, any STT, LLM, TTS, or realtime model, and deployment to LiveKit Cloud or self-hosted servers.

Murf
murf.ai
Murf provides text-to-speech, speech-to-speech, and dubbing through APIs and SDKs, including the Falcon text-to-speech model built for voice agents, alongside voice generation tools for creators and businesses.
OpenAI Realtime API
OpenAI · developers.openai.com
The OpenAI Realtime API streams audio to and from OpenAI's realtime speech models for voice agents, connecting over WebRTC, WebSocket, or SIP, and it handles turns, tools, interruptions, and handoffs with the OpenAI Agents SDK.
Pipecat
Daily · pipecat.ai
Pipecat is an open-source Python framework for real-time voice and multimodal agents, maintained by Daily, that composes speech, language, and voice services into pipelines over WebRTC, SIP, or PSTN and runs self-hosted or on Pipecat Cloud.
Retell AI
retellai.com
Retell AI is a platform for phone-call voice agents with a visual flow builder, testing and simulation, call analytics, SIP and existing phone number support, human handoff, and on-premises deployment for regulated industries.
Rime
rime.ai
Rime builds voice models for real-time conversation, with the Coda and Mist text-to-speech models for cascaded voice agents, the Prism speech-to-speech model, speech-to-text, and enterprise voice cloning.

Soniox
soniox.com
Soniox provides one speech AI API for real-time speech-to-text, text-to-speech, and speech translation across many languages, with low-latency streaming and support for speakers who switch languages mid-conversation.

Speechmatics
speechmatics.com
Speechmatics builds speech-to-text APIs for real-time and batch transcription across many languages, accents, and multiple speakers, aimed at regulated industries, with deployment on-device, on-premises, or in the cloud.

Synthflow
synthflow.ai
Synthflow provides AI voice agents for contact centers, BPOs, and multi-location brands that answer and handle phone calls inside the phone systems and CRMs those businesses already run.

TEN Framework
theten.ai
TEN Framework is a framework for real-time multimodal voice agents supported by Agora, with RTC and WebSocket transport and voice activity and turn detection, released under Apache-2.0 with added conditions that bar competing with Agora's offerings.

Twilio ConversationRelay
Twilio · twilio.com
Twilio ConversationRelay connects phone calls on Twilio to the AI model of your choice through a WebSocket API, handling speech-to-text, text-to-speech, and automatic language detection so applications exchange text with the call.
Ultravox
Inworld · ultravox.ai
Ultravox is a real-time voice AI platform built on speech-native models, with APIs, a console, telephony integrations, and tools for building voice agents; the company was formerly known as Fixie and is now part of Inworld.
Inworld acquired Ultravox in 2026, and ultravox.ai remains available for building voice agents.
Vapi
vapi.ai
Vapi is a developer platform for voice AI agents that make and receive phone and web calls, configured through an API, dashboard, or config-as-code, with a choice of model, transcription, and voice providers, squads of assistants, and simulations.
How to choose
- 01Pipeline type
- Cascaded pipelines chain speech-to-text, a language model, and text-to-speech for control over each stage, while speech-to-speech models reply directly in audio with fewer hops.
- 02Latency and turn-taking
- Streaming, voice activity and end-of-turn detection, and interruption handling decide whether a conversation feels natural.
- 03Languages and voices
- Supported languages, accents, language switching within a call, and voice cloning options differ between providers.
- 04Telephony and transport
- WebRTC, WebSocket, SIP, and phone number support determine whether an agent can run in a browser, an app, or a call center.
- 05Deployment and compliance
- On-premises or self-hosted models, data retention terms, and certifications such as HIPAA and SOC 2 matter for regulated calls.
- 06Control
- Code-first frameworks give full control over the pipeline, while hosted platforms trade some flexibility for dashboards, call flows, and built-in analytics.
Frequently asked questions
What is a voice AI agent?
- A voice AI agent is software that holds a spoken conversation and takes actions for a user, such as answering a support call or booking an appointment. It either chains speech-to-text, a language model, and text-to-speech, or uses a single speech-to-speech model, and it calls tools during the conversation. Frameworks such as LiveKit Agents and Pipecat and platforms such as Vapi are used to build them.
LiveKit vs Pipecat: what is the difference?
- Both are open-source frameworks for real-time voice agents that let developers mix speech-to-text, language model, and text-to-speech providers in code. LiveKit Agents runs agents as participants in LiveKit's WebRTC rooms, offers Python and Node.js SDKs, and deploys to LiveKit Cloud or self-hosted servers. Pipecat, maintained by Daily, is a Python pipeline framework that works over WebRTC, SIP, or phone transports and runs self-hosted or on Pipecat Cloud.
Vapi vs Retell: how do they compare?
- Vapi and Retell AI are both hosted platforms for voice agents that handle phone calls. Vapi is API-first, configured through its API, dashboard, or config-as-code files, with a choice of model, transcription, and voice providers and squads of assistants. Retell AI centers on a visual flow builder, call testing and simulation, analytics, and human handoff, and it offers on-premises deployment. Both connect to telephony providers such as Twilio.
Deepgram vs ElevenLabs: what is the difference?
- Deepgram offers speech-to-text, text-to-speech, and a Voice Agent API for enterprise voice applications, with real-time streaming transcription. ElevenLabs offers text-to-speech, voice cloning, speech-to-text, sound effects, and a platform for conversational voice agents, with a large library of voices. Many teams combine providers, for example Deepgram for transcription and ElevenLabs for synthesis, through a framework such as Pipecat or LiveKit Agents.
Related guides
AI agents hubOther categories
- Model providers
- Inference providers
- GPU clouds
- Serving engines
- Vector databases
- Embedding and reranking models
- Document parsing and data ingestion
- Fine-tuning and post-training
- Agent frameworks
- Agent infrastructure
- Web search APIs
- MCP platforms
- AI gateways
- LLM observability and evaluation
- Guardrails and AI security
- AI coding tools
Product details change quickly; check each vendor’s site for current features, licensing and pricing.