# Fish Audio review (2026): What it is and why enterprises choose Bland Speech

By Samuel Oram-Jones

Fish Audio bills text to speech by the byte, so its $15 price for the s2.1-pro model is $15 per million UTF-8 bytes, the unit computers use to store text. An English letter takes one byte, and characters in some other scripts take more, which means the price per character depends on the language a team writes in.

For English that price sits well under Bland Speech's $40 per million characters, and the catches are elsewhere. As of November 2026, Fish Audio's free model carries no latency or data processing guarantees, its open-weight licence covers commercial use only while a company earns under $500,000 a year, and its newest model, Drama 3, costs $50 per million bytes at list price.

## What is Fish Audio and how does it work?

Fish Audio is a voice AI platform from Hanabi AI Inc. that sells text to speech through a web app, a REST API, and SDKs, and Fish Audio says it is "Trusted by 8M+ users." Fish Audio says its models run in production at HeyGen, Retell, and Sanas. The company announced $52 million in seed funding on its first anniversary, and on 10 November 2026 it introduced Live 3, a full-duplex voice model, which means a model that keeps listening while it speaks.

### Which models does Fish Audio sell?

Fish Audio runs several models side by side. The s2.1-pro model is the one Fish Audio recommends for production, and s2.1-pro-free is the same model at $0 under fair-use limits. Drama 3, launched on 13 October 2026, is built for voice direction and content creation. The s2-pro model, the previous generation, was open-sourced on 9 March 2026, and the older s1 retires on 31 December 2026, after which s2.1-pro serves its requests.

A request whose model header is missing or misspelled falls back to s2.1-pro, so a call meant for Drama 3 that carries a typo is served by the cheaper model without an error.

### How fast is it, and how does it stream?

Fish Audio streams two ways: HTTP streaming for text a developer already has, and a WebSocket for text that arrives word by word, as a language model writes it. Latency is stated in three different ways. Fish Audio says s2-pro delivers 100ms to first audio, which is the time between a request and the first sound arriving, and it also quotes a latency of under 70 ms for enterprise. For s2.1-pro, the production model, Fish Audio promises "improved quality, latency, and throughput" over s2-pro.

Bland publishes 199ms to first audio. Each of those numbers is the company's own, so none of them is a race, and Fish Audio's are lower.

### How do voices and cloning work?

An instant clone needs at least 10 seconds of clean audio, and a minute or two of single-speaker speech improves fidelity. A professional clone runs only in the web app on a paid plan, from at least 10 minutes of audio, and Fish Audio puts the whole flow at 1 to 2 hours. Fish Audio requires a confirmation that the voice is the user's own or that the speaker consented, and a released professional voice can earn a share when other creators pay to use it. Fish Audio states 83 languages for s2.1-pro.

### What is it like to build on?

Fish Audio ships a Python SDK, a JavaScript SDK, an MCP server, and REST and WebSocket APIs, and publishes a migration guide for teams moving from ElevenLabs or OpenAI. Output formats for text to speech are mp3 by default, wav, pcm, and opus, with a default of 44,100 Hz. Phone networks carry mu-law, an 8-bit audio encoding at 8 kHz, so audio from those formats is converted before a caller hears it. Fish Audio's agent platform dials US, Canada, and Japan destinations and imports SIP numbers, which is a separate product from the text to speech API.

## How much does Fish Audio cost?

Fish Audio's API is pay as you go, with no subscription fee or monthly minimum. Prices run per million UTF-8 bytes: s2.1-pro, s2-pro, and s1 are $15.00 each, Drama 3 is $50.00 at list price, and s2.1-pro-free is $0.00. Drama 3 carries a launch price of $10.00 per million bytes through 30 November 2026, which is 80% off list. Fish Audio converts a million bytes to about 180,000 English words, or about 12 hours of speech.

Concurrency, the number of requests running at once, rises with prepaid spend: 5 requests under $100 paid, 15 at $100 or more, and 50 at $1,000 or more. The web app has its own plans, priced in pounds: Free gives 8,000 credits a month for personal, non-commercial use, App Plus costs £11.49 a month at list price with 250,000 credits, and App Pro costs £76.99 a month with 2,000,000 credits.

Self-hosting starts from a $10K one-time platform fee, with a $10K monthly minimum spend and a 12-month commitment. The Fish Audio Community License Agreement v1.0, dated 13 October 2026, makes the open weights free for research and non-commercial use, and commercial use stays free only while annual revenue is under USD 500,000. Hosted services and larger commercial use need a separate licence.

Bland Speech is $0.04 per 1,000 characters, which is $40 per million, pay as you go with no monthly plan and 50,000 free characters once for a new account. For English text, s2.1-pro costs less than half of that. Drama 3 at list price costs more.

## Which sounds more human, Fish Audio or Bland Speech?

Bland Speech v3 ranked first of 13 AI voices in [Design Arena's Audio Realism Bench results](https://notes.designarena.ai/audio-realism-benchmark-measuring-realism-in-audio-models/) of 4 August 2026, at 1365, while the human recording scored 1500, and Fish Audio wasn't in those results. In that test, vetted native speakers hear two recordings of one script and pick the voice that sounds more like a real person.

Fish Audio titled its Drama 3 launch post "Fish Audio Drama 3: The Most Emotive Text-to-Speech Model We've Built." On control, Fish Audio is looser than Bland Speech. Its S2 models take natural-language tags in square brackets that are "not limited to a fixed set," and Drama 3 takes plain-language direction, sounds, pauses, and delivery limited to specific words. Bland Speech takes a set list of performance tags, such as [laughs], [sighs], and [say warmly], which Bland says is the set the model was trained on, plus two numeric controls for expressiveness and stability and timed pauses from 0.1 to 10 seconds.

A team that wants free-form stage directions will find that room in Fish Audio, and a team that wants a voice that passes for a person on a phone line has Bland Speech's dated blind result to go on.

## Is Fish Audio secure enough for enterprise calls?

Fish Audio's enterprise terms are strongest on deployment and thinner on audited certification. Fish Audio says a SOC 2 audit is underway and that the report will go to customers under NDA once it is complete. Fish Audio says it supports HIPAA-aligned configurations and can sign a BAA, a business associate agreement, for qualifying healthcare workloads. Zero Data Retention, which means request payloads are never persisted, is available on enterprise contracts, and independent penetration testing runs as part of an ongoing compliance program.

By default data stays in the United States, hosted on Google Cloud, with inference from US and Asia-Pacific (Tokyo) edge regions. The self-hosted tier runs in a VPC, on-premise, air-gapped, or in a sovereign cloud, and the enterprise uptime SLA is 99%.

Bland, the company, is SOC 2 Type II audited, HIPAA compliant with a signed BAA, and FedRAMP 20x Class C certified at Moderate, granted on 17 November 2026, and Bland says speech synthesis runs inside that FedRAMP boundary. Zero-retention processing is available for sensitive workflows. Enterprise customers get dedicated orchestration and dedicated GPUs run by Bland, and Enterprise pricing is custom, set with Bland's sales team.

## Is Fish Audio worth it for enterprises?

Fish Audio is worth it for a team that wants a low list price per English character, free-form direction of delivery, open weights to run itself, and a self-hosted tier. Fish Audio is a weaker fit for a team that needs a SOC 2 report in hand before launch.

| | Bland Speech | Fish Audio |
|:--|:--|:--|
| Price | $0.04 per 1,000 characters ($40 per million), pay as you go | s2.1-pro $15 per million UTF-8 bytes, Drama 3 $50 at list |
| Monthly plan | None | None for the API, app plans from £11.49 |
| Free to start | 50,000 characters once | s2.1-pro-free at $0 under fair-use limits, without latency or DPA guarantees |
| Latency, as published | 199ms to first audio | 100ms for s2-pro, and under 70 ms as Fish Audio's enterprise figure |
| Phone-ready output | mu-law at 8 kHz, PCM at 8 to 48 kHz | mp3, wav, pcm, and opus |
| Instant clone | One clean sample of about 10 seconds | At least 10 seconds |
| Professional clone | 30 minutes or more, with consent and rights review | At least 10 minutes, web app only |

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

- **Price:** Fish Audio meters bytes and Bland Speech meters characters, which match for English text and differ for scripts that take several bytes per character.
- **Free to start:** Fish Audio states that the free model has no time-to-first-audio or data processing agreement guarantees.
- **Latency:** Time to first audio is the gap between a request and the first sound. Both columns are as each company publishes it, with different methods and models.
- **Phone-ready output:** Mu-law is the 8 kHz encoding phone networks carry. Fish Audio's text to speech formats are the four listed.
- **Professional clone:** A professional clone is trained on a person's recordings. Bland says every clone is private to a customer's organization and that the customer confirms the voice is theirs to use.

</details>

### Fish Audio pros and cons

<div class="cmp-proscons" data-rival="fish-audio"><div class="cmp-proscons-pros"><p><strong>Pros</strong></p><ul><li>A list price of $15 per million UTF-8 bytes on the production model</li><li>A free API model, s2.1-pro-free, for development</li><li>Natural-language bracket tags that are not limited to a fixed set</li><li>83 languages stated for s2.1-pro</li><li>An open-weight s2-pro, and an enterprise tier that runs on a customer's own infrastructure</li></ul></div><div class="cmp-proscons-cons"><p><strong>Cons</strong></p><ul><li>The free model has no latency or data processing guarantees</li><li>Drama 3 costs $50 per million bytes at list price</li><li>The open-weight licence covers commercial use only below USD 500,000 of annual revenue</li><li>Professional cloning runs only in the web app</li><li>A SOC 2 audit is underway, with no report yet</li></ul></div></div>

### Why Bland Speech is the Fish Audio alternative for phone calls

Bland Speech was built for phone calls, and the v3 model ranked first of 13 AI voices in Design Arena's results of 4 August 2026. Bland Speech outputs mu-law at 8 kHz, so audio reaches a caller without a conversion step, and its WebSocket takes text as a language model writes it, with a new context interrupting a turn that is playing, which is how barge-in works when a caller talks over the voice.

Pipecat has a first-party Bland service and LiveKit Agents has a Python plugin, and an OpenAI-compatible endpoint lets a team switch a client by changing a base URL, a key, and a model. One standard rate per character applies, with no byte arithmetic, and Bland says its model and speech are its own rather than a wrapper around a third party.

<div class="cmp-proscons cmp-proscons-own"><div class="cmp-proscons-pros"><p><strong>Pros</strong></p><ul><li>First of 13 AI voices in Design Arena's results of 4 August 2026, at 1365</li><li>199ms to first audio, as Bland publishes it</li><li>Mu-law at 8 kHz and PCM from 8 to 48 kHz for phone audio</li><li>A WebSocket for streamed text with barge-in</li><li>One standard rate per character, with no monthly plan</li><li>First-party Pipecat service and a LiveKit Agents plugin</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>

### Verdict: Fish Audio or Bland Speech for phone calls?

Bland Speech is the pick for a phone line or a voice agent, because it has a dated blind-listening result, phone-ready audio, a published 199ms to first audio, and Bland's SOC 2 Type II audit and FedRAMP 20x Class C certification at Moderate. Fish Audio is the cheaper English API and the looser creative tool, and a team that mostly narrates content or runs its own servers has good reason to use it. A [Bland Speech vs Fish Audio comparison](/compare/bland-speech-vs-fish-audio) sets the two side by side, [Fish Audio alternatives](/blog/fish-audio-alternatives) covers the wider field, and Bland Speech's [text to speech API](/text-to-speech-api) has the endpoints. More reviews sit in the [text to speech comparison hub](/compare/text-to-speech) and its [review index](/compare/text-to-speech/reviews), and a team can [try Bland Speech](/speech) with 50,000 free characters.

## Frequently asked questions

### Is Fish Audio free?

Fish Audio has a free API model, s2.1-pro-free, which costs $0 under fair-use limits and carries no latency or data processing guarantees. The web app's Free plan gives 8,000 credits a month and limits generated audio to personal, non-commercial use. Bland Speech gives a new account 50,000 free characters once, then bills $0.04 per 1,000 characters.

### How much does Fish Audio cost per character?

Fish Audio prices by UTF-8 byte: $15 per million bytes for s2.1-pro, which is about $15 per million characters for English text, and $50 per million bytes for Drama 3 at list price. Characters in other scripts take more than one byte, so the same text costs more. Bland Speech is $40 per million characters.

### Is Fish Audio open source?

Fish Audio open-sourced s2-pro on 9 March 2026, and its Community License Agreement v1.0 makes commercial use free only while annual revenue is under USD 500,000.

### Does Fish Audio work for phone calls?

Fish Audio's agent platform dials US, Canada, and Japan destinations, and its text to speech API returns mp3, wav, pcm, or opus. A team feeding a phone network has to convert to mu-law at 8 kHz, which Bland Speech returns directly.

### How does Fish Audio voice cloning compare with Bland Speech?

Both clone an instant voice from a clip of at least 10 seconds, and both require consent for a professional clone. Fish Audio runs professional cloning in its web app from 10 minutes of audio, while Bland Speech trains a professional clone on 30 minutes or more after a consent and rights review. A separate [cloning comparison](/compare/bland-speech-vs-fish-audio-voice-cloning) goes through the details.

### What is the best Fish Audio alternative for a voice agent?

Bland Speech is the alternative for a voice agent that answers phone calls, because Bland Speech v3 ranked first of 13 AI voices in Design Arena's results of 4 August 2026, Bland publishes 199ms to first audio, and the API returns mu-law. A team ranking low-latency options can read best low-latency text to speech APIs.
