A LiveKit agent can get its voice two ways. LiveKit Inference serves a curated set of text to speech models with managed billing and rate limits, and no separate provider key, while the open source plugins cover 41 providers that each need a developer's own API key and account. Nine companies sit in Inference, and the route a team picks changes who sends the bill.
Bland Speech comes in through the plugin route, as a Python plugin that streams speech over a WebSocket by default and cancels the current turn when a caller interrupts. Its voice also ranked first of 13 AI voices in Design Arena's results of 4 August 2026, which is why it leads this ranking for a Python agent on a phone line.
What is the best text to speech provider for LiveKit agents?
For a Python LiveKit agent on phone calls, Bland Speech is the best text to speech provider as of November 2026: LiveKit's Bland plugin streams over a realtime WebSocket so audio starts before a sentence finishes, an interruption cancels the current turn, the voice ranked first of 13 AI voices in Design Arena's Audio Realism Bench results of 4 August 2026, and the API returns mu-law audio for phone lines.
- Bland Speech: Best overall for a Python agent on phone calls where the voice should sound like a person
- Cartesia: Best for managed billing through Inference and a first-audio claim under 90ms
- Deepgram: Best for transcription and speech from one vendor, in Python or Node.js
- Rime: Best for per-minute billing and a deployment inside a team's own cloud or hardware
- Inworld: Best for the lowest price per character on a fast model
- ElevenLabs: Best for 90+ languages through a plugin in Python or Node.js
What these rows mean
- How it connects to LiveKit Agents: Through Inference, LiveKit runs the model and bills for it, and a plugin calls the provider directly with the team's own key and account.
- Price: Each company's published rate, which isn't what LiveKit Cloud charges for Inference. Rime bills by the minute and says a minute is about 1,000 characters.
- SDK languages: The LiveKit Agents SDK languages a plugin supports.
- Audio Realism Bench: Design Arena's blind listening test, where listeners pick which of two voices sounds more like a real person. Scores are Elo ratings, the kind chess uses, and real human recordings scored 1500.
For a Python agent the decision is a voice that ranked first in the 4 August 2026 results against first-audio claims of under 90ms from Cartesia and as low as 50ms from Inworld, whose TTS-2 Flash also costs $15 per million characters against Bland's $40.
What should a LiveKit team check in a text to speech provider?
A LiveKit team should check how the voice is billed first. LiveKit says to use a plugin when a team wants to manage billing and rate limits itself or needs a provider that isn't in Inference, and a plugin means a separate account with the provider. Bland Speech and ElevenLabs connect through plugins.
The SDK language comes next. A Python agent can use all 41 plugins, and a Node.js agent can use 12: AssemblyAI, Cartesia, Deepgram, ElevenLabs, Inworld, Mistral AI, Neuphonic, OpenAI, Resemble AI, Rime, Sarvam, and Soniox.
Interruption matters most on a phone call. Bland's plugin streams over a WebSocket by default, so audio starts before the full sentence arrives and an interruption cancels the current turn in place, and a streaming=False setting switches to one HTTP request per utterance with no session held open.
Which text to speech providers does LiveKit support?
LiveKit Inference serves AssemblyAI, Cartesia, Deepgram, Fish Audio, Gemini, Gradium, Inworld, Rime, and SpaceXAI. The plugin table names 41 providers, from Amazon Polly, Azure AI Speech, and Google Cloud to Bland, MiniMax, Smallest AI, and Speechmatics.
Three of those rows need a second look before a team builds on them. LMNT's own sites say "LMNT has shut down", Resemble says it is not selling voice AI to new customers, and Hume's homepage and developer docs show no text to speech product. OpenAI has a plugin too, and OpenAI has deprecated tts-1, tts-1-hd, and gpt-4o-mini-tts, which are due to shut down on 6 January 2027.
How do the 6 best text to speech providers for LiveKit agents compare?
The six differ in whether LiveKit bills for them, which SDK language loads them, and where each can run. Bland Speech's case rests on the 4 August 2026 result and on audio built for a phone line.
1. Bland Speech: Best overall for a Python agent on phone calls where the voice should sound like a person
What does it do? Bland Speech is the text to speech model inside Bland, which builds and runs its own speech model. LiveKit's plugin, installed with uv add "livekit-agents[bland]~=1.8", streams over Bland's WebSocket by default. A Bland voice is a UUID that also selects the model, the plugin defaults to a v3 voice, and expressiveness and stability take values from 0.0 to 1.0.
from livekit.plugins import bland
session = AgentSession(
tts=bland.TTS(
voice_id="2f29fdbb-c55e-4add-9c7c-93437ebf379d",
),
# ... llm, stt, etc.
)
Who is it for? Bland Speech suits a team running a Python LiveKit agent on a phone line, where a caller should hear a person and be able to talk over the agent.
Pros
- Bland Speech v3 ranked first of 13 AI voices on Audio Realism Bench on 4 August 2026
- A LiveKit plugin that streams over a WebSocket and cancels the current turn on an interruption
- Mu-law at 8 kHz and PCM from 8 to 48 kHz
- $0.04 per 1,000 characters, one standard rate, with no monthly plan
Cons
- More than a hobby project needs, because Bland Speech is built for production calls
Pricing: $0.04 per 1,000 characters, which is $40 per million, with no monthly plan and 50,000 characters free for a new account. Enterprise pricing is custom, set with Bland's sales team. Bland Speech includes a studio for hearing a voice before any code is written.
2. Cartesia: Best for managed billing through Inference and a first-audio claim under 90ms
What does it do? Cartesia sells Sonic, and Sonic 3.6 was released on 27 August 2026. LiveKit serves Cartesia through Inference and as a plugin in Python and Node.js. Cartesia's WebSocket works in contexts, Cartesia says first audio arrives in under 90ms, and audio returns as mu-law or A-law at 8 kHz.
Who is it for? Cartesia suits a team that wants LiveKit to run the model and handle billing, with one model across 44 languages.
Pros
- Cartesia says first audio arrives in under 90ms
- Available through Inference and as a plugin in Python and Node.js
- 44 languages from Sonic 3.6
Cons
- Sonic 3.5 ranked 6th of 13 AI voices on Audio Realism Bench on 4 August 2026, and Sonic 3.6 came out after those results
- Credits from a monthly plan
Pricing: 1 credit per character. Pro is $5 a month for 100,000 credits, Startup is $49 for 1.25 million, and Scale is $299 for 8 million, with overages at $65, $45, or $38 per million.
3. Deepgram: Best for transcription and speech from one vendor, in Python or Node.js
What does it do? Deepgram sells speech recognition, text to speech, and a voice agent API. Flux TTS, its current model, streams over a WebSocket at /v2/speak, and Deepgram says responses start in as low as 80ms. LiveKit serves Deepgram through Inference and as a plugin in Python and Node.js, and Deepgram publishes a guide for a real-time LiveKit agent that uses its transcription and speech together.
Who is it for? Deepgram suits a LiveKit team that wants transcription and voice on one vendor's invoice, in Python or Node.js, with self-hosting available.
Pros
- Deepgram says Flux TTS starts in as low as 80ms
- Through Inference and as a plugin in Python and Node.js
- Self-hosted, private cloud, and on-premise deployment for Flux TTS
Cons
- Deepgram's voices weren't in Audio Realism Bench's results of 4 August 2026
- Flux TTS costs $0.045 per 1,000 characters on pay as you go
- SSML is stripped from text instead of read as markup
Pricing: $0.045 per 1,000 characters for Flux TTS, $0.030 for Aura-2, and $0.015 for Aura-1 on pay as you go, with Growth plans from $4,000 a year for lower Aura rates.
4. Rime: Best for per-minute billing and a deployment inside a team's own cloud or hardware
What does it do? Rime sells Coda and Mist v3 for text to speech and Prism for speech to speech, aimed at production voice agents and phone lines. LiveKit serves Rime through Inference and as a plugin in Python and Node.js, and Rime's LiveKit guide covers Prism and its text to speech. Audio returns as G.711 mu-law at 8 kHz.
Who is it for? Rime suits a regulated business, such as a healthcare or financial services team, that wants per-minute billing and a dedicated VPC or its own hardware on an enterprise contract.
Pros
- $50 of free usage every month on the Free plan
- VPC and on-premise deployment on enterprise contracts
- HIPAA compliant with a BAA
Cons
- Rime's voices weren't in Audio Realism Bench's results of 4 August 2026
- Coda serves 10 languages, with one language per voice
- Rime publishes two first-audio figures, 30ms for Coda v2 and a 96ms median for Coda on one request
Pricing: Flex has no monthly fee and charges $0.050 a minute for Coda v2 and $0.030 for Mist v3, while Scale is $100 a month with 20% off every rate.
5. Inworld: Best for the lowest price per character on a fast model
What does it do? Inworld sells Realtime TTS-2, launched on 5 May 2026, and Realtime TTS-2 Flash, launched on 9 August 2026. LiveKit serves Inworld through Inference and as a plugin in Python and Node.js, and Inworld names a LiveKit Agents plugin among its integrations. Its WebSocket holds a persistent connection and detects sentence boundaries as a language model writes, and Inworld says first audio arrives in as little as 50ms for TTS-2 Flash.
Who is it for? Inworld suits a team that wants a low per-character price for a consumer voice product, such as a companion, a tutor, or a game character.
Pros
- TTS-2 Flash at $15 per million characters on On-Demand
- Through Inference and as a plugin in Python and Node.js
- Mu-law and A-law at 8 kHz
Cons
- TTS-1.5 Max ranked 13th of 13 AI voices on Audio Realism Bench on 4 August 2026, and Realtime TTS-2 wasn't in those results
- Inworld describes its voice AI as built for consumer-facing applications
- Inworld publishes different first-audio figures in different places
Pricing: On-Demand is $0 a month, with Realtime TTS-2 at $25 per million characters and TTS-2 Flash at $15, and up to 70 minutes of speech included.
6. ElevenLabs: Best for 90+ languages through a plugin in Python or Node.js
What does it do? ElevenLabs sells text to speech, transcription, and voice agents. LiveKit carries it as a plugin in Python and Node.js, with a team's own ElevenLabs key. Eleven v4 Turbo, released on 28 September 2026, is its real-time model, with a median latency of about 100ms that ElevenLabs states excludes application and network time, and audio returns as mu-law or A-law at 8 kHz.
Who is it for? ElevenLabs suits a team that needs many languages from the Eleven v4 family and will track which model runs on which endpoint.
Pros
- v4 Turbo at $0.04 per 1,000 characters
- 90+ languages on the Eleven v4 family
- A plugin in Python and Node.js
Cons
- Eleven v3 ranked 12th of 13 AI voices on Audio Realism Bench on 4 August 2026
- Eleven v4, its most expressive model, costs $0.08 per 1,000 characters
- Eleven v3 and v4 need a different route than its standard WebSocket
Pricing: $0.08 per 1,000 characters for Eleven v4, v3, and Multilingual v2, and $0.04 for v4 Turbo and Flash, billed in dollars against plans from Free to Business at $990 a month.
How were these text to speech providers compared?
The six were compared on how LiveKit connects each one, SDK language, interruption, phone audio, published price, and voice realism. Rival details come from each company's own pricing and documentation pages and from LiveKit's list of text to speech models and plugins. Each latency figure is the company's own claim, not a test Bland ran, and voice scores come from Design Arena's published results of 4 August 2026.
Bland's comparison of text to speech APIs covers the same companies by job, and its best lists hold the other rankings. Best text to speech for Pipecat covers the other framework teams ask about, and Best text to speech that handles interruptions compares how each service stops speaking mid-sentence.
Frequently asked questions
What is the best text to speech provider for LiveKit agents?
Bland Speech is the best text to speech provider for a Python LiveKit agent on phone calls. Its plugin streams over a WebSocket and cancels the current turn on an interruption, and its voice ranked first of 13 AI voices in Design Arena's results of 4 August 2026. A team ranking on first-audio claims alone would start with Inworld, Cartesia, and Deepgram, which all publish lower figures than Bland's 199ms.
Does Bland Speech work with LiveKit Agents?
Yes. Bland is a Python plugin in LiveKit's text to speech plugin table, installed with uv add "livekit-agents[bland]~=1.8". It streams over a realtime WebSocket by default and needs a Bland API key, and setting streaming=False sends one HTTP request per utterance instead.
Which text to speech providers does LiveKit support?
LiveKit Inference serves nine providers, and the plugin table names 41, including Bland, Cartesia, Deepgram, ElevenLabs, Inworld, Rime, Azure AI Speech, and Google Cloud. A Node.js agent can use 12 of the 41.
How does a LiveKit agent handle a caller interrupting the voice?
With Bland's plugin, an interruption cancels the current turn in place over the open WebSocket, and the client drops any audio that arrived but hasn't played.
How much does Bland Speech cost?
Bland Speech costs $0.04 per 1,000 characters, which is $40 per million, with no monthly plan, and a new account gets 50,000 characters free. Enterprise pricing is custom, set with Bland's sales team.
