# Bland Speech vs Kokoro (2026): Is a free open-weight model or a hosted API better for phone calls?

Kokoro is an 82-million-parameter model that anyone can download under the Apache 2.0 licence, and its author says the model has "been deployed in numerous projects and commercial APIs". A team that picks it for a phone agent gets the weights and a Python library, and then builds a server, a streaming layer, and phone audio on top.

Kokoro comes from an individual author, not a company, and the author publishes the weights and the library while others run the commercial APIs. The question is whether a free model is worth that work when a hosted API charges by the character and runs the machine.

## Which is better, Bland Speech or Kokoro?

Bland Speech is the better choice for production phone calls as of November 2026, because Bland runs the model and a team using Kokoro runs it alone. A team calls an endpoint and pays $0.04 per 1,000 characters, and Bland Speech v3 ranked first of 13 AI voices in Design Arena's Audio Realism Bench results of 4 August 2026, where Kokoro wasn't included.

Kokoro makes more sense for a team that wants an Apache 2.0 model on its own hardware, at no licence cost, and will own every fix. A team weighing more open models can start with [open-source text to speech compared](/blog/open-source-tts-compared).

<div class="cmp-usewhen"><div class="cmp-usewhen-own"><p><strong>Use Bland Speech when</strong></p><ul><li>Callers on a phone line should hear a voice that sounds like a person</li><li>A team wants a text to speech API with a printed price and a company running the model</li><li>An existing OpenAI audio speech integration should move by changing a base URL, a key, and a model name</li><li>A script needs performed delivery, such as a laugh or a warm read, written inline as tags</li><li>A team wants a private voice clone from one 10-second sample, with consent confirmed for every voice</li></ul></div><div><p><strong>Use Kokoro when</strong></p><ul><li>A team will host an 82-million-parameter model on its own hardware under the Apache 2.0 licence</li><li>A team will write the streaming server, the scaling, and the monitoring around a Python library</li><li>A product can work with one of Kokoro's 54 fixed voices and will test how each language sounds</li><li>A team will buy Kokoro through a third-party host and accept that host's terms</li></ul></div></div>

| Feature | Bland Speech | Kokoro |
|---|---|---|
| Status | Sold as a studio and an API, run by Bland | Open-weight model and Python library from an individual author |
| Audio Realism Bench, 4 August 2026 | Ranked 1st of 13 AI voices, 1365 | Not in those results |
| Price | $0.04 per 1,000 characters ($40 per million) | Free weights under Apache 2.0, plus a team's own compute |
| Streaming | HTTP chunks, and a WebSocket that takes text as a language model writes it | Python library that returns audio chunks from a generator |
| Expressive control | Inline tags such as [laughs] and [say warmly], plus expressiveness and stability settings | A speed setting, voice choice, and IPA pronunciation in the text |
| Phone audio | Mu-law at 8 kHz, and PCM from 8 to 48 kHz | Author's example plays audio at 24 kHz |
| Voices | Library voices, plus instant clones from about 10 seconds and professional clones from 30 minutes or more | 54 fixed voices across 8 languages in version 1.0 |
| Compliance | Bland is SOC 2 Type II audited, HIPAA compliant with a signed BAA, and FedRAMP 20x Class C certified at Moderate | Self-hosted weights run on a team's own infrastructure, which carries the audit |

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

- **Status:** Kokoro's author is an individual, credited as @rzvzn, and warns that any website with "kokoro" in its domain isn't affiliated with the model page or its author.
- **Audio Realism Bench, 4 August 2026:** Design Arena's blind listening test, where listeners hear two clips of the same line from two models and pick the one that sounds more like a real person. A higher score means listeners picked that voice more often.
- **Price:** Bland Speech costs $0.04 per 1,000 characters in the studio or the API. Kokoro's weights cost nothing to download, and a team pays for the machine and the engineers that keep it running.
- **Streaming:** The author's quickstart runs a pipeline that returns audio in a generator, one chunk per text split. Bland Speech streams over POST /v2/tts and over a WebSocket at WSS /v2/tts/ws.

</details>

## How do Bland Speech and Kokoro compare?

The biggest difference is who stands behind the model. Bland runs Bland Speech and sells it by the character, and Kokoro's author publishes weights and a library for anyone to use.

### Bland Speech overview

Bland Speech is the text to speech model inside Bland, a voice AI provider that runs its own model and speech rather than wrapping third parties, and it is a closed-source model Bland built itself. Bland sells it on its own as a studio and an API on the [Bland Speech](/speech) page.

A team sends text over HTTP to POST /v2/tts and gets raw audio or a WAV file back, or streams language model output into a WebSocket, where a new context interrupts the turn that is playing. POST /v2/audio/speech follows OpenAI's audio speech contract, and the [text to speech API](/text-to-speech-api) page has the code. Tags in square brackets, such as [laughs], [sighs], or [say warmly], tell the model how to deliver the words after them. Bland publishes 199ms to first audio.

<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>Bland runs the model, so a team has no GPUs to size or patch</li><li>One standard rate, $0.04 per 1,000 characters, with no monthly plan</li><li>An OpenAI-compatible endpoint, so an existing integration moves by changing a base URL, a key, and a model name</li><li>Mu-law at 8 kHz and PCM from 8 to 48 kHz, ready for phone calls</li><li>A first-party Pipecat service and a Python plugin for LiveKit Agents</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>

### Kokoro overview

Kokoro is an open-weight text to speech model with 82 million parameters, built on StyleTTS 2 and ISTFTNet as a decoder-only model, and its author says it delivers quality comparable to larger models while being significantly faster and more cost-efficient. Version 1.0 came out on 27 January 2025, after version 0.19 on 25 December 2024.

The code is on GitHub as `hexgrad/kokoro`, the weights are on Hugging Face as `hexgrad/Kokoro-82M`, and both are Apache 2.0. The Hugging Face listing shows more than 13 million downloads, and the GitHub repository's last commit is dated 6 August 2025. The library installs with `pip install kokoro`, and the repository includes a JavaScript port folder.

<div class="cmp-proscons" data-rival="kokoro"><div class="cmp-proscons-pros"><p><strong>Pros</strong></p><ul><li>Apache 2.0 weights, free to download and deploy</li><li>An 82-million-parameter model, small beside larger text to speech models</li><li>54 voices across 8 languages in version 1.0</li><li>Custom pronunciation by writing IPA into the text</li><li>Deployed in numerous projects and commercial APIs, per its author</li></ul></div><div class="cmp-proscons-cons"><p><strong>Cons</strong></p><ul><li>An individual author, not a company, stands behind the model</li><li>Hosting, streaming, and scaling belong to the team that runs it</li><li>The author notes weakness on short utterances, especially under 10 to 20 tokens</li><li>Non-English support may be thin, per the author</li><li>Two CC BY training datasets need attribution</li></ul></div></div>

### What does running Kokoro take?

Running Kokoro takes a Python environment, hardware, and a service a team writes itself. The author's quickstart installs the `kokoro` package and runs a `KPipeline` that hands back audio chunks from a generator, one per text split. A team building a voice agent decides how text gets in, how audio reaches a phone carrier, and what happens under load.

Kokoro's author warns of a limit that matters for short replies: "Weakness on short utterances, especially less than 10-20 tokens." Design Arena says the phone-agent prompts in its own test are single turns that tend to be shorter than its other prompts.

With Bland Speech the hardware is Bland's. Enterprise customers get dedicated orchestration and dedicated GPUs run by Bland.

### How do Bland Speech and Kokoro handle voices and languages?

Kokoro ships voices as fixed packs, and Bland Speech adds clones. Kokoro version 1.0 has 54 voices across American English, British English, Japanese, Mandarin Chinese, Spanish, French, Hindi, Italian, and Brazilian Portuguese, and its author says support for non-English languages may be absent or thin because of weak grapheme-to-phoneme conversion or a lack of training data.

Bland Speech makes an instant clone from one clean sample of about 10 seconds. A professional clone is fine-tuned on 30 minutes or more of audio and passes Bland's consent and rights review, and every clone is private to the customer's organization, with the customer confirming the voice is theirs to use.

### How do Bland Speech and Kokoro handle phone audio and security?

Bland Speech returns mu-law at 8 kHz, the 8-bit format built for telephony, or PCM at 8, 16, 24, 44.1, or 48 kHz. Kokoro's example plays its output at 24,000 Hz, so how audio reaches a phone line is the team's decision.

Bland is SOC 2 Type II audited, HIPAA compliant with a signed BAA, and FedRAMP 20x Class C certified at Moderate, granted 17 November 2026, and Bland's trust page says speech synthesis runs inside the FedRAMP boundary. Kokoro's weights run on a team's own infrastructure, so that team carries the audit.

## How much do Bland Speech and Kokoro cost?

Kokoro costs nothing to download and something to run, and Kokoro served by third-party hosts costs far less per character than Bland Speech. Kokoro's author says that in April 2025 the market rate for Kokoro served over an API is under $1 per million characters of text input, against $40 per million on Bland Speech.

### Bland Speech pricing

Bland Speech costs $0.04 per 1,000 characters, in the studio or the API, with no monthly plan. New accounts get 50,000 characters free, and credits don't expire. A 60-second note costs about $0.04, a 10-minute narration about $0.36, and a million characters $40.

Enterprise pricing is custom, set with Bland's sales team.

### Kokoro pricing

The weights are free under Apache 2.0, and Kokoro's author says the model "can be deployed anywhere from production environments to personal projects". A team that hosts Kokoro pays for the machine under it and the engineering hours to keep it running. A team that buys Kokoro from a third-party host pays that host's rate, which the author put under $1 per million characters in April 2025, and accepts that host's terms.

### What does a month of speech cost?

On Bland Speech, 100,000 characters a month, about 2 hours of audio, costs $4, 1 million characters costs $40, and 10 million costs $400. Kokoro at the author's April 2025 market rate would cost under $0.10, under $1, and under $10 for the same volumes. What a team pays Bland for is the model running, the voice, and the streaming layer, and whether that is worth the difference depends on what an engineer's time costs.

## Which sounds more human, Bland Speech or Kokoro?

Kokoro's author calls its quality comparable to larger models, and Kokoro wasn't in Design Arena's Audio Realism Bench results of 4 August 2026, where Bland Speech v3 ranked first of 13 AI voices at 1365 and real human recordings scored 1500.

Listeners in that test hear two clips of the same line, blind to which model made which, and pick the one that sounds more like a real person. The author's claim is a statement and not a score. Design Arena publishes the method and the scores in its [Audio Realism Bench article](https://notes.designarena.ai/audio-realism-benchmark-measuring-realism-in-audio-models/).

## Why is Bland Speech the best Kokoro alternative?

Kokoro saves money on licensing and spends it on engineering time, and Bland Speech takes that work back. Bland runs the model, the rate is one standard price, a new account starts with 50,000 free characters, and the voice ranked first of 13 AI voices on Audio Realism Bench on 4 August 2026.

Teams still weighing open models can read [Kokoro vs Chatterbox](/blog/kokoro-vs-chatterbox) and the [Bland Speech vs Chatterbox](/compare/bland-speech-vs-chatterbox) comparison, or Bland's [comparison of text to speech APIs](/compare/text-to-speech) and its [head-to-heads](/compare/text-to-speech/head-to-heads) for more vendors.

### How was this comparison made?

The criteria are status, hosting, price, streaming, voices, phone audio, and security. Kokoro details come from its author's model card, voice notes, and GitHub repository, and voice scores come from Design Arena's published results of 4 August 2026.

## Frequently asked questions

### Is Kokoro free?

The weights are, under the Apache 2.0 licence. A team still pays for the hardware that runs the model, or pays a third-party host. Bland Speech costs $0.04 per 1,000 characters, which is $40 per million, with 50,000 characters free for a new account.

### Does Kokoro have a hosted API?

Kokoro's author publishes weights and a Python library, and says Kokoro has been deployed in "numerous projects and commercial APIs" run by others. The author also warns that websites with "kokoro" in the domain aren't affiliated with the model page or its author.

### Is Bland Speech better than Kokoro?

For phone calls and voice agents, yes. Bland Speech v3 ranked first of 13 AI voices on Audio Realism Bench on 4 August 2026, and Kokoro wasn't in those results. Kokoro suits teams that want free weights under Apache 2.0 and will run them.

### How many languages does Kokoro support?

Kokoro version 1.0 covers 8 languages and 54 voices, counting American and British English as one language. Its author says support for non-English languages may be absent or thin.

### What voices does Kokoro come with?

Kokoro version 1.0 ships 54 voice packs, from 20 American English voices down to one French voice. Bland Speech makes an instant clone from about 10 seconds of audio and a professional clone from 30 minutes or more, with consent confirmed for every voice.

### Does Kokoro work on phone calls?

Kokoro's example plays audio at 24,000 Hz, and a team decides how audio reaches a phone line. Bland Speech returns mu-law at 8 kHz as well as PCM from 8 to 48 kHz, so audio goes onto a phone line without conversion.

### Is Kokoro still maintained?

The GitHub repository's last commit is dated 6 August 2025, and the Hugging Face model card was last modified on 10 April 2025. Bland Speech is a current Bland product, sold on its own since 11 March 2026.
