# Text to speech vs speech-to-speech for voice agents (2026): What a phone call needs from a voice

By Samuel Oram-Jones

A woman calls to report a tree that came down in a storm, and halfway through reading out her address she stops to answer her son. An agent that only receives a transcript gets the address, the aside, and her answer to the child as one run of words, and has to work out which of them was meant for it. Bland opens its case for Tofu-1 on this scene to question how most voice agents are built.

Bland says most voice systems in use today are cascades: speech recognition turns the caller's audio into text, a language model writes a reply, and text to speech (TTS) turns that reply back into audio. Speech-to-speech systems take audio in and produce audio out from one model. Bland sells one of each, Bland Speech as TTS in production and Tofu-1 as a speech-to-speech research preview, as of November 2026.

## What is the difference between text to speech and speech-to-speech?

Text to speech turns finished text into spoken audio and does nothing else, while speech-to-speech takes a caller's audio and produces the reply's audio from one system, so what the sound carries reaches the model along with the words.

In a cascade, TTS is the last stage and never hears the call. Each stage hands the next one a smaller version of what came before, and Bland says each handoff "discards part of the call: who was speaking, who they were speaking to, whether they hesitated, and whether they had finished." A transcript spells every word right and still can't say who a word was for.

Tofu-1 is built the other way. Bland describes it as full duplex, which means it listens and speaks at the same time, so a caller who says "mm-hmm" and a caller who says "no, it's 4, not 9" are heard as two different things while the agent is still mid-sentence. Bland also says Tofu-1 keeps working on a request while it talks and never says a task is done before it is.

## What does a cascade lose, and what does it keep?

A cascade loses whatever the transcript can't hold, and it keeps a set of separate parts that a team can price, swap, and read.

Bland describes one test for the loss, which it calls the TV Test. An agent asks a caller to confirm she wants to cancel a policy, she is still hunting for her policy number, and a television in the room says "Yes!" A transcript-based agent can take the television's "yes" as her answer, while Tofu-1 is built to attribute the "yes" to the television and ask the caller to confirm. Bland labels the scenario illustrative, not a recorded model output, and adds a limit: if the background makes a caller impossible to understand, Tofu-1 should say so and ask for a repeat instead of guessing.

The parts a cascade keeps are real. A TTS voice sits behind its own endpoint, so a team can change the voice without touching the language model, and the words the voice speaks exist as text before they are spoken. TTS also bills in a unit that is easy to read: Bland Speech charges $0.04 per 1,000 characters, pay as you go, with 50,000 characters free once for a new account. OpenAI bills its realtime models in audio tokens instead, $10 per million input tokens and $20 per million output tokens for gpt-realtime-2.1-mini and $32 and $64 for gpt-realtime-2.1. Characters and audio tokens don't convert into each other, so a team compares the two bills by running calls and reading the invoice.

A cascade can also be interrupted. Bland Speech's WebSocket, a connection that stays open so text and audio travel both ways, stops the turn that is playing the moment a new context arrives, and Bland calls interruption and barge-in "a normal part of the protocol". Barge-in is a caller talking over an agent. What the voice can't do is judge whether that interruption was agreement or a correction, because it receives text and not the call.

## Who is moving from text to speech to speech-to-speech?

OpenAI is moving its customers, xAI (now SpaceXSI) and ElevenLabs sell both kinds side by side, and Bland offers a TTS in production and a speech-to-speech research preview.

| Company | Text to speech | Speech-to-speech |
|---|---|---|
| Bland | Bland Speech v3, $0.04 per 1,000 characters | Tofu-1, a research preview with limited access |
| OpenAI | tts-1, tts-1-hd, and gpt-4o-mini-tts, deprecated, shutting down 6 January 2027 | gpt-realtime-2.1 and gpt-realtime-2.1-mini on the Realtime API, billed in audio tokens |
| xAI Grok | $15 per million characters for Grok TTS | grok-voice-3.0, added in October 2026 |
| ElevenLabs | $0.08 per 1,000 characters for Eleven v4, v3, and Multilingual v2, and $0.04 for v4 Turbo and Flash | Realtime v1 models, marked OpenAI Realtime-compatible |

<details><summary>What these rows mean</summary>

- **Text to speech:** each company's published model names and list price for turning text into audio.
- **Speech-to-speech:** the model each company sells that takes audio in and returns audio.
- **Billing units:** characters for TTS, audio tokens for OpenAI's realtime models, so no row says one costs less than another.

</details>

OpenAI's case is the one with a date on it. Its tts-1, tts-1-hd, and gpt-4o-mini-tts models shut down on 6 January 2027, and OpenAI points customers to gpt-realtime-2.1-mini on its Realtime API, which it calls "a change of API workflow" because the speech endpoint doesn't accept that model. A team that wants to keep a cascade has a smaller move. Bland Speech's POST /v2/audio/speech implements OpenAI's /audio/speech contract, so an existing integration changes its base URL, API key, and model name. Bland's guide to [migrating from OpenAI TTS](/blog/migrate-from-openai-tts) covers the steps.

## Which should a voice agent use, text to speech or speech-to-speech?

A voice agent that must go live on a schedule should run on text to speech such as Bland Speech, and a team with crowded, noisy, or interruption-heavy calls should also test speech-to-speech such as Tofu-1 in its research preview.

Bland says Tofu-1 is meant for calls where conditions are worst and mistakes cost most: calls from moving cars and hospital rooms, over weak signal, with several people talking at once. Bland says access is "open to a limited number of enterprises", and it is looking for teams in banking, healthcare, insurance, and government that want to test where the model works and where it breaks. Requests go through Bland's [Tofu-1 page](/speech-to-speech).

<div class="cmp-usewhen"><div class="cmp-usewhen-own"><p><strong>Use text to speech, such as Bland Speech, when</strong></p><ul><li>A team will assemble its own cascade, choosing a speech recognition service and a language model and sending each reply to a voice over a WebSocket</li><li>A team will move off OpenAI's text to speech models before 6 January 2027 and keep its OpenAI SDK</li><li>A team will put a voice on a phone line and needs 8 kHz mu-law output</li></ul></div><div><p><strong>Use speech-to-speech, such as Tofu-1, when</strong></p><ul><li>A team will test a model that hears audio directly on calls with background noise, several speakers, or a television on</li><li>A team will run calls where a caller's mid-sentence correction has to change what the agent says next</li><li>A team in banking, healthcare, insurance, or government will join a research preview and report where it breaks</li></ul></div></div>

How a voice sounds is a separate question from how a system listens. Design Arena's [Audio Realism Bench](https://notes.designarena.ai/audio-realism-benchmark-measuring-realism-in-audio-models/), a blind listening test in which native speakers hear two recordings of the same line and pick the one that sounds more like a real person, ranked Bland Speech v3 first of 13 AI voices at 1365 on 4 August 2026, and OpenAI's GPT Realtime 2, one of the realtime models on OpenAI's audio pricing, 4th at 1119. The test compares how voices sound and can't show how well a model follows a call, so a high score from either kind of system says nothing about the TV Test.

<div class="cmp-proscons cmp-proscons-own"><div class="cmp-proscons-pros"><p><strong>Pros</strong></p><ul><li>Bland Speech v3 ranked first of 13 AI voices on Audio Realism Bench on 4 August 2026</li><li>One standard rate, $0.04 per 1,000 characters, with no monthly plan</li><li>A WebSocket that takes text as a language model writes it and lets a caller interrupt</li><li>8 kHz mu-law and PCM from 8 to 48 kHz for phone lines</li><li>An OpenAI-compatible endpoint for teams leaving OpenAI's TTS models</li></ul></div><div class="cmp-proscons-cons"><p><strong>Cons</strong></p><ul><li>More than a hobby project needs, because Bland Speech is built for production calls</li></ul></div></div>

## How does Bland Speech work inside a voice agent cascade?

Bland Speech sits at the end of a cascade as one endpoint: POST /v2/tts for complete text and WSS /v2/tts/ws for text that arrives in pieces, such as tokens from a language model.

Bland publishes 199ms to first audio, the wait between sending text and the first sound, counted to the first audio chunk of a streamed request and not to the finished file. Output comes as PCM from 8 to 48 kHz or as mu-law at 8 kHz, the 8-bit format built for telephone audio. Pipecat, a framework for building voice agents, has a first-party Bland service maintained by the Pipecat team, and LiveKit Agents has a Bland plugin for Python. Bland documents both endpoints with its [text to speech API](/text-to-speech-api), and Bland's guides on [the best TTS for voice agents](/blog/best-tts-for-voice-agents) and [the best TTS for LLM streaming](/blog/best-tts-for-llm-streaming) go deeper on the cascade side.

Bland's [comparison of text to speech providers](/compare/text-to-speech) and its [developer guides](/compare/text-to-speech/guides) cover the rest, and [Bland Speech](/speech) has a studio for hearing a voice.

## Frequently asked questions

### What is the difference between text to speech and speech-to-speech?

Text to speech turns written text into audio, and speech-to-speech takes a caller's audio and returns audio from one system. In a voice agent cascade, text to speech is the last of three stages after speech recognition and a language model. Speech-to-speech skips the text in the middle, so who spoke, who was addressed, and whether a caller hesitated stay available to the model.

### Is speech-to-speech better than text to speech for a voice agent?

Not across the board. Bland's argument for speech-to-speech is about calls with background noise, overlapping voices, and interruptions, and Tofu-1 is a research preview with limited access. Text to speech is the part of a voice agent with a published price, an endpoint, and a first-audio figure today, and Bland Speech is Bland's version of it.

### What is Tofu-1?

Tofu-1 is Bland's full duplex speech-to-speech research preview, built for calls from moving cars and hospital rooms, over weak signal, with several people talking at once. Bland says it understands audio, listens while it speaks, and keeps working while it talks. Requests for the research preview go through Bland's [Tofu-1 page](/speech-to-speech).

### Can a text to speech voice be interrupted mid-sentence?

Yes. On Bland Speech's WebSocket, a new context_id immediately preempts the turn that is playing, so a caller talking over an agent ends the sentence the agent was speaking. Deciding that the caller meant to interrupt is the job of the recognition and language model stages.

### Is OpenAI shutting down its text to speech models?

Yes. OpenAI has deprecated tts-1, tts-1-hd, and gpt-4o-mini-tts, and they shut down on 6 January 2027. OpenAI's replacement is gpt-realtime-2.1-mini on its Realtime API, a different API workflow. Bland Speech's POST /v2/audio/speech implements the same contract as OpenAI's speech endpoint, so a team can change a base URL, key, and model name instead.

### How is speech-to-speech priced compared with text to speech?

In different units. Bland Speech charges $0.04 per 1,000 characters, Grok TTS charges $15 per million characters, and ElevenLabs charges $0.04 or $0.08 per 1,000 characters depending on the model. OpenAI prices its realtime models in audio tokens, gpt-realtime-2.1-mini at $10 per million input and $20 per million output. A team compares the two by running a set of calls.

### Which voice agent frameworks work with Bland Speech?

Pipecat has a first-party Bland service with classes for the WebSocket and the HTTP endpoint, and LiveKit Agents has a Bland plugin for Python. Bland Speech also has an OpenAI-compatible endpoint, and its native endpoints are POST /v2/tts and WSS /v2/tts/ws.
