# How to migrate from OpenAI TTS to Bland Speech (2026): Three changed values, and what doesn't carry over

By Samuel Oram-Jones

Moving an application off OpenAI's text to speech models can be a three-value edit. Bland Speech accepts OpenAI's audio speech request shape at `/v2/audio/speech`, so a team that calls the endpoint through an OpenAI SDK changes a base URL, an API key, and a model name, and leaves the rest of its code where it is.

The edit is small and the differences sit in the fine print. Fifteen OpenAI voice names collapse onto three Bland voices, an `instructions` prompt is ignored, and `speed` only accepts 1.0, so each difference is worth knowing before a call line depends on it.

## Why are teams moving off OpenAI's text to speech endpoint?

As of November 2026, OpenAI will shut down tts-1, tts-1-hd, and gpt-4o-mini-tts on 6 January 2027, and the replacement OpenAI names, gpt-realtime-2.1-mini, isn't accepted by the `/v1/audio/speech` endpoint that all three share. OpenAI calls the move "a change of API workflow", because the replacement runs over OpenAI's Realtime API, which connects over WebRTC in a browser or WebSocket on a server, instead of a single HTTP request.

A team that wants to keep one request in and audio out has two paths: rebuild around the Realtime API, or send the same request shape to another provider. Bland describes the second path as changing the base URL, API key, and model, because `/v2/audio/speech` implements OpenAI's `/audio/speech` contract. A team that already runs its whole conversation through OpenAI's Realtime API has no speech endpoint to swap, so its decision is a different one.

## What maps to what when moving from OpenAI TTS to Bland Speech?

A request keeps its shape: `model`, `input`, `voice`, and `response_format` carry over, the model name becomes `btts-3`, and each OpenAI voice name resolves to one of three Bland voices.

| What a team touches | Bland Speech | OpenAI TTS |
| --- | --- | --- |
| Endpoint | POST /v2/audio/speech, same request and error shapes | POST /v1/audio/speech |
| Model | btts-3, or btts-2 for an older voice; OpenAI model ids are not aliased | tts-1, tts-1-hd, and gpt-4o-mini-tts, shutting down 6 January 2027 |
| Voice | OpenAI names mapped to Karen, Matthew, or River, or any Bland voice UUID | Built-in names such as alloy, coral, and nova |
| Output formats | mp3, opus, aac, flac, wav, and pcm at 24 kHz; mu-law at 8 kHz on /v2/tts | mp3, opus, aac, flac, wav, and pcm at 24 kHz, with no mu-law |
| Delivery | `instructions` accepted and ignored, `speed` 1.0 only; tags, pause markers, and two controls on /v2/tts | `instructions` prompt on gpt-4o-mini-tts |
| Streaming | Audio chunks over HTTP, SSE returns 400, WebSocket at /v2/tts/ws for LLM tokens | HTTP chunked streaming, SSE not on tts-1 or tts-1-hd, WebSocket on the Realtime API |
| Price | $0.04 per 1,000 characters | tts-1 $15 per million characters, tts-1-hd $30, gpt-4o-mini-tts per token |

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

- **Endpoint:** Bland's compatible endpoint takes OpenAI's request fields and answers with OpenAI's error envelope, so an SDK's error handling keeps working. An unknown model returns `model_not_found`, and a model that doesn't match the chosen voice returns `model_voice_mismatch`.
- **Voice:** Bland's native endpoints take only a voice UUID. The compatible endpoint also accepts OpenAI's documented voice names and maps each to a Bland voice.
- **Output formats:** Bland's compatible endpoint returns mp3, opus, aac, flac, wav, and pcm, and Bland's native `/v2/tts` and WebSocket offer `mulaw` fixed at 8,000 Hz. OpenAI's `/v1/audio/speech` pcm is 24 kHz.
- **Streaming:** OpenAI's replacement path is the Realtime API. Bland's WebSocket takes text as a language model writes it and cuts off the turn playing when a new `context_id` arrives.
- **Price:** each is the published rate. Bland's is one standard rate for studio and API, and OpenAI's gpt-4o-mini-tts is $0.60 per million text input tokens and $12.00 per million audio output tokens.

</details>

Bland maps OpenAI names this way: alloy, ash, coral, and fable go to River, amber, august, lily, nova, and shimmer go to Karen, and ballad, blue, echo, onyx, sage, and verse go to Matthew. A team that gave two characters two different OpenAI voices can find both now speaking in one Bland voice, and a voice UUID from Bland's voice list opens the full catalog.

Delivery is the one real rewrite. gpt-4o-mini-tts takes a natural-language `instructions` prompt, with "Speak in a cheerful and positive tone." as OpenAI's own example. Bland accepts that field and ignores it, so a prompt that steered tone on OpenAI does nothing on Bland. Steering moves into the text: bracket tags such as `[say warmly]` or `[sighs]` placed before the words they shape, pause markers such as `<|0.8|>` with a length from 0.1 to 10.0 seconds, and two controls called `expressiveness` and `stability`, each from 0.0 to 1.0, on `/v2/tts`. Bland publishes the tag set verbatim and says invented variants are ignored or garbled, and the tags work on current-generation voices only.

## How do I migrate from OpenAI TTS to Bland Speech, step by step?

A switch takes six steps: change three values, rebuild the voice table, move delivery into the text, choose the endpoint for phone audio, stream for voice agents, and run both providers on a slice of calls before cutover.

1. A team changes the base URL to `https://api.bland.ai/v2`, the API key to a Bland key, and the model to `btts-3`:

```python
import os
from pathlib import Path

from openai import OpenAI

client = OpenAI(
    api_key=os.environ["BLAND_API_KEY"],
    base_url="https://api.bland.ai/v2",
)

with client.audio.speech.with_streaming_response.create(
    model="btts-3",
    voice="coral",
    input="Hello from Bland.",
) as response:
    response.stream_to_file(Path("hello.mp3"))
```

2. A team builds a voice table with each OpenAI voice name in one column and the Bland voice it resolves to beside it. A GET to `/v1/voices` returns every voice a team can use, and Speech Studio renders a line with the same models the API uses, so candidates can be auditioned there first.
3. A team rewrites steering prompts as tags, pause markers, and controls, and tests each against a recording of the old output.
4. A team that sends audio to a phone line moves that traffic to `/v2/tts`, which takes `{"audio": {"encoding": "mulaw", "sample_rate": 8000}}` and returns telephony audio without a conversion step. A team that converts OpenAI's 24 kHz PCM into mu-law for a call line can drop that conversion.
5. A voice agent that waits for a complete language model reply before asking for speech adds the model's whole generation time to every answer, and Bland recommends the WebSocket for new conversational applications. One socket serves one conversation, and Bland returns audio as binary frames until it sends `utterance_end` for the turn:

```json
{ "type": "init", "voice": "29158307-9893-4149-8a75-bc9ce313d64e", "audio": { "encoding": "mulaw", "sample_rate": 8000 } }
{ "type": "speak", "context_id": "turn-42", "text": "Hello" }
{ "type": "speak", "context_id": "turn-42", "text": ", how can I help?" }
{ "type": "end_of_turn", "context_id": "turn-42" }
```

6. A team routes a small share of calls to Bland, compares recordings against OpenAI's output for the same lines, and moves the rest after those recordings pass a blind listen.

Bland publishes 199ms to first audio on a streamed request, and Bland says that figure runs to the first audio chunk rather than the finished file. Text goes out as the model writes it, with its leading spaces and punctuation intact, because Bland picks the synthesis boundaries itself and client-side buffering adds latency.

## What does the move cost, and what can't come along?

A move to Bland Speech costs more per character than tts-1 or tts-1-hd, and no OpenAI custom voice comes along, because Bland voices are made on Bland from audio.

### What does a million characters cost on each side?

tts-1 is $15 per million characters and tts-1-hd is $30, against Bland Speech's $40, so at 10 million characters a month the list prices come to $150, $300, and $400. tts-1 itself shuts down on 6 January 2027. gpt-4o-mini-tts and the Realtime replacement are priced per token, which leaves no per-character figure to set beside Bland's.

Bland Speech is pay as you go with no monthly plan, 50,000 characters are free once for a new account, and each request carries a $0.001 minimum, so a 12-character confirmation bills the minimum. Billing follows delivery: a synthesis that fails before any bytes are written isn't charged. Enterprise pricing is custom, set with Bland's sales team.

### What can't move from OpenAI TTS to Bland Speech?

OpenAI limits custom voices made from recordings to eligible customers, allows at most 20 per organization, takes samples of 60 seconds or less, and requires a consent recording in an approved phrase. None of that carries over. Bland voice IDs can't be reused across providers, so a team clones each voice again from its original recording, with the speaker's permission.

On Bland, an instant clone needs one clean sample of about 10 seconds, a professional clone is fine-tuned on 30 minutes or more of audio and goes through Bland's consent and rights review, every clone is private to the customer's organization, and the customer confirms the voice is theirs to use.

## Will callers hear a difference after moving from OpenAI TTS to Bland Speech?

Bland Speech v3 ranked first of 13 AI voices in Design Arena's [Audio Realism Bench results of 4 August 2026](https://notes.designarena.ai/audio-realism-benchmark-measuring-realism-in-audio-models/), at 1365, with the human recording at 1500, while OpenAI's GPT-4o Mini TTS ranked 9th at 1011 and GPT Realtime 2 ranked 4th at 1119.

Neither tts-1 nor tts-1-hd was in those results, and GPT Realtime 2 belongs to the Realtime family, which isn't one of the three speech endpoint models on a shutdown date. In the benchmark, native speakers pick blind which of two clips of the same text sounds more like a real person, and it has a Phone Agents category. A board scores a voice across hundreds of prompts and can't say how one team's lines will sound, so the sturdier test is to render a call script's most repeated lines through both providers and have a listener who can't see the labels choose.

A head-to-head sits at [Bland Speech vs OpenAI TTS](/compare/bland-speech-vs-openai-tts), and [OpenAI TTS alternatives](/blog/openai-tts-alternatives) covers the other places a team can go. Moves from other providers have their own [text to speech migration guides](/compare/text-to-speech/guides), the [text to speech comparison hub](/compare/text-to-speech) holds every comparison, the [text to speech API](/text-to-speech-api) page lists each endpoint, and [Bland Speech](/speech) has the studio.

## Frequently asked questions

### Is Bland Speech a drop-in replacement for OpenAI's audio speech endpoint?

For basic synthesis, yes: `/v2/audio/speech` implements OpenAI's `/audio/speech` contract, so a team changes the base URL, API key, and model name, and keeps its SDK. The exceptions are that OpenAI model ids such as tts-1 aren't aliased, `speed` accepts only 1.0, `instructions` is ignored, and the `sse` stream format returns a 400.

### Which OpenAI voice becomes which Bland voice?

Bland maps alloy, ash, coral, and fable to River, amber, august, lily, nova, and shimmer to Karen, and ballad, blue, echo, onyx, sage, and verse to Matthew. Names are case-insensitive, an unrecognized name returns a 400, and a Bland voice UUID reaches the full catalog.

### What happens to OpenAI's instructions field and speed setting?

Bland accepts `instructions` and ignores it, and accepts `speed` only at 1.0, returning a 400 for any other value. A team that steered tone with instructions rewrites that steering as bracket tags, pause markers, and the two controls on `/v2/tts`.

### Can Bland Speech produce mu-law audio for a phone line?

Yes, on Bland's native endpoints: `mulaw` at 8,000 Hz on `/v2/tts` and `/v2/tts/ws`. The OpenAI-compatible endpoint returns mp3, opus, aac, flac, wav, and pcm, so phone traffic goes to the native endpoint.

### Does a Pipecat or LiveKit agent need code changes to move?

A small change in each. Pipecat supports OpenAI and Bland as text to speech services, and the Pipecat team maintains Bland's, installed with `uv add "pipecat-ai[bland]"`. LiveKit Agents has OpenAI plugins for Python and Node.js and a Bland plugin for Python, installed with `uv add "livekit-agents[bland]~=1.8"`.

### Can an OpenAI custom voice move to Bland Speech?

No. A voice is tied to the provider that made it, so a team clones it again on Bland from the original recording. An instant clone needs one clean sample of about 10 seconds, and a professional clone needs 30 minutes or more and passes Bland's consent and rights review.

### Is Bland Speech cheaper than OpenAI TTS?

No, not against tts-1 at $15 per million characters or tts-1-hd at $30. Bland Speech is $40 per million characters, pay as you go with no monthly plan, and gpt-4o-mini-tts is priced per token, so no per-character comparison exists for it.
