Fish Audio bills speech at $15 per million UTF-8 bytes and MiniMax bills it at $100 per million characters, and a byte is not a character. English text runs about one byte per letter while Chinese, Hindi, and other scripts run more, so a price quoted per byte and a price quoted per character only line up for English.
Both companies sell expressive speech with inline tags and voice cloning, which leaves billing, audio formats, and cloning rules as the real differences. A team putting either one on a phone line should also weigh Bland Speech, whose v3 model ranked first of 13 AI voices in Design Arena's Audio Realism Bench results of 4 August 2026, the results where MiniMax's Speech-02 HD ranked 7th and Fish Audio wasn't entered.
Which is better, Fish Audio or MiniMax?
Fish Audio is the better pick for a team that wants the lower list price and 83 languages on its production model, and MiniMax is the better pick for a team that needs mu-law phone audio and a WebSocket that takes text a character at a time, as of November 2026. Fish Audio's recommended production model, s2.1-pro, costs $15 per million UTF-8 bytes, while MiniMax's speech-2.8-turbo costs $60 per million characters and speech-2.8-hd $100.
Fish Audio fits when
- A script will be directed in plain language on Drama 3, with sounds and pauses placed where they happen
- A team will read a community licence before shipping open weights commercially, since Fish Audio's is free only while annual revenue stays under $500,000
- A brand will train a professional voice clone in Fish Audio's web app from 10 minutes or more of audio
MiniMax fits when
- A team will pipe a language model's output into a bidirectional WebSocket and cancel a turn with a task_cancel message
- A phone integration will request mu-law at 8 kHz as pcmu_raw or pcmu_wav
- A team will buy speech by the character on pay as you go or as monthly audio packages from $5 to $999
What these rows mean
- Audio Realism Bench, 4 August 2026: Design Arena's blind listening test, where listeners hear two clips of the same line and pick the one that sounds more like a real person. Each score belongs to the model named in the cell, and models released after those results are not in them.
- Price: Fish Audio bills per million UTF-8 bytes, a count of stored bytes, and MiniMax and Bland Speech bill per character.
- First audio: The pause between sending text and hearing the first sound. Fish Audio quotes "<70 ms" latency for enterprise customers, and gives s2-pro a time to first audio of 100ms. Bland publishes 199ms to first audio.
- Phone audio: Mu-law is the 8-bit format US phone networks carry. Fish Audio's documented formats are mp3, wav, pcm, and opus, and a phone integration converts them to mu-law.
- Instant voice clone: Fish Audio says a minute or two of clear speech improves fidelity.
What are the catches with Fish Audio and MiniMax?
Fish Audio's catches are byte billing, a free model without production guarantees, and a licence with a revenue ceiling on its open weights, while MiniMax's catches are a higher per-character rate and cloned voices that disappear after 7 days unused.
Fish Audio
Fish Audio calls s2.1-pro its recommended production model and a "Production option for workloads that need TTFA and DPA guarantees", where TTFA is time to first audio. The same model is offered as s2.1-pro-free at $0 under fair-use limits, and Fish Audio says it comes without those guarantees. Drama 3 launched on 13 October 2026 for voice direction and costs $50 per million UTF-8 bytes. A missing or misspelled model header falls back to s2.1-pro, so a typo returns s2.1-pro speech instead of Drama 3 speech.
Professional Voice Clone runs only in the web app, on the Plus or Pro plan, from at least 10 minutes of clean audio, and requires the customer to confirm permission to clone the voice.
On security, Fish Audio says a SOC 2 audit "is currently underway" with the report available under NDA when complete, and that it supports "HIPAA-aligned configurations" and can sign a BAA for qualifying healthcare workloads. Data stays in the United States by default, on Google Cloud, and zero data retention is available on enterprise contracts. Fish Audio open-sourced S2 on 9 March 2026, and the S2-Pro model card on Hugging Face names a research licence. The fish-speech repository carries a Community License Agreement dated 13 October 2026, which allows commercial use free only while annual revenue is under USD 500,000 and asks for a separate licence for hosted services and larger commercial use.
Pros
- $15 per million UTF-8 bytes on s2.1-pro, the recommended production model
- 83 languages on s2.1-pro
- Inline direction written in natural language, not limited to a fixed set of tags
- Open S2 weights, published in March 2026
Cons
- Billed in UTF-8 bytes, not characters
- Documented output formats are mp3, wav, pcm, and opus
- SOC 2 audit underway, with the report due under NDA
- The free model carries no TTFA or DPA guarantees
- Commercial use of open weights is free only under $500,000 in annual revenue
MiniMax
MiniMax sells speech next to its language, video, image, and music models. MiniMax's current speech models are speech-2.8-hd and speech-2.8-turbo, released on 23 January 2026, with speech-2.6 and speech-02 still sold as legacy models.
The speech-2.8 models take 19 inline interjection tags, among them laughs, sighs, gasps, and breaths, and pause markers from 0.01 to 99.99 seconds. Pronunciation can be overridden inline with Pinyin, IPA (the International Phonetic Alphabet), or Jyutping, and the whisper emotion works on the 2.6 models but not on speech-2.8.
Rapid voice cloning needs a source clip of 10 seconds to 5 minutes and costs $1.50 per voice, charged when the voice first speaks. MiniMax deletes a cloned voice that isn't used within 7 days.
On streaming, the HTTP endpoint returns server-sent events, a one-way stream from server to client. A bidirectional WebSocket accepts text "at any granularity (even character by character)" and buffers it into sentences on the server. Speech-2.6 and speech-2.8 cover 40 languages, and speech-02 covers 24.
Pros
- 19 inline interjection tags on speech-2.8 models
- A bidirectional WebSocket that takes text a character at a time, with a cancel message
- Mu-law at 8 kHz for telephony
- Voice design from a written description, at $3 a voice
Cons
- $60 and $100 per million characters on the current models
- Cloned voices are deleted after 7 days without use
- Cloning is billed at $1.50 a voice on first use
- The whisper emotion isn't supported on speech-2.8
What do Fish Audio, MiniMax, and Bland Speech cost?
Fish Audio's s2.1-pro costs $15 per million UTF-8 bytes, MiniMax's speech-2.8-turbo $60 per million characters, MiniMax's speech-2.8-hd $100, and Bland Speech $40 per million characters ($0.04 per 1,000).
Fish Audio pricing
Fish Audio says "There are no subscription fees or monthly minimums for API access." Billing is per million UTF-8 bytes: $15 for s2.1-pro, s2-pro, and s1, $50 for Drama 3 (a launch price of $10 applies through 30 November 2026), and $0 for the free model. Fish Audio converts a million bytes to about 180,000 English words, or about 12 hours of speech. The web app is priced apart from the API, and its Free plan covers personal, non-commercial use only.
MiniMax pricing
MiniMax charges $100 per million characters for speech-2.8-hd and $60 for speech-2.8-turbo on pay as you go, and the legacy speech-2.6 and speech-02 models carry the same pair of rates. Monthly audio packages, which MiniMax says support all speech models, cost $5 for 100,000 audio points, $30 for 300,000, $99 for 1,100,000, $249 for 3,300,000, and $999 for 20,000,000, with 10% off for three months and 20% off for a year.
Bland Speech pricing
Bland Speech costs $0.04 per 1,000 characters, pay as you go, with no monthly plan, at one standard rate in the studio and the API whether or not a script is directed. A new account gets 50,000 free characters once, and credits don't expire. Each request carries a minimum charge of $0.001, and some public-library voices add a per-character creator fee.
What a month costs on each
At 1 million characters of English a month, Fish Audio's s2.1-pro costs about $15, MiniMax's speech-2.8-turbo $60, speech-2.8-hd $100, and Bland Speech $40. At 10 million characters the same rates come to about $150, $600, $1,000, and $400. None of those figures include a phone line or taxes, and Fish Audio's s2.1-pro is the cheapest of the four on list price.
Which sounds more human, Fish Audio, MiniMax, or Bland Speech?
In Design Arena's Audio Realism Bench results of 4 August 2026, MiniMax's Speech-02 HD ranked 7th of 13 AI voices at 1093, Fish Audio wasn't in those results, and Bland Speech v3 ranked first at 1365 against a human recording at 1500.
- #1BlandBland Speech v31365
- 2ndMicrosoftMAI-Voice-21224
- 3rdxAIGrok TTS1156
- 4thOpenAIGPT Realtime 21119
- 5thGoogleGemini 2.5 Pro TTS1109
- 7thMiniMaxSpeech-02 HD1093
Ranks are among 13 AI voices, and MiniMax's own voice is on the list. Real human recordings scored 1500, at the right edge of this scale. Bars start at 600, not zero. Source: Design Arena's published results.
In the test, listeners hear two clips of one line, blind to which model made which, and choose the one that sounds more like a real person. Scores are Elo ratings, the chess-style number that goes up each time a voice is chosen over another, and phone agents are one of the prompt categories.
Each score belongs to the model that was entered. MiniMax's speech-2.8 models weren't in that chart, and Fish Audio's Drama 3 launched on 13 October 2026, after those results. Design Arena explains its method in its Audio Realism Bench article.
Why do people comparing Fish Audio and MiniMax choose Bland Speech?
Bland Speech bills by the character at one standard rate, so a team weighing Fish Audio's bytes against MiniMax's characters and audio points gets a third price that needs no conversion, from the voice that ranked first in the 4 August 2026 results, Bland Speech v3. Fish Audio's s2.1-pro list rate is lower than Bland Speech's $40, and Bland publishes 199ms to first audio, slower than the 100ms Fish Audio publishes for s2-pro.
Bland Speech's WebSocket at /v2/tts/ws takes text as a language model writes it, and a new context cuts off the turn that is playing, which is how a caller who talks over an agent gets heard. Output is mu-law at 8 kHz or PCM from 8 to 48 kHz, and an endpoint that follows OpenAI's speech contract lets a team switch by changing a base URL, a key, and a model name. The text to speech API page has the endpoints.
On security, Bland is SOC 2 Type II audited and HIPAA compliant with a signed BAA, where Fish Audio says its SOC 2 audit is underway, and Bland is FedRAMP 20x Class C certified at Moderate, granted on 17 November 2026. Bland's trust page says speech synthesis runs inside the FedRAMP boundary. Where Fish Audio publishes open S2 weights, Bland Speech is a closed-source model Bland built itself. An instant clone needs one clean sample of about 10 seconds, and every clone is private to the customer's organization, with the customer confirming the voice is theirs to use.
Pros
- Bland Speech v3 ranked first of 13 AI voices on Audio Realism Bench on 4 August 2026
- One standard rate of $0.04 per 1,000 characters, with no byte counting, audio points, or monthly plan
- Mu-law and PCM for phone calls, with interruption built into the WebSocket
- An OpenAI-compatible endpoint, a Pipecat service, and a LiveKit Agents plugin
Cons
- More than a one-off narration project needs, because Bland Speech is built for production calls
More pairings sit on the rival comparisons page, and the text to speech comparison hub covers every vendor. Bland has pages on Bland Speech against MiniMax and Bland Speech against Fish Audio, plus lists of Fish Audio alternatives and MiniMax speech alternatives. The Bland Speech page has the studio and the price.
Frequently asked questions
Is Fish Audio better than MiniMax?
Fish Audio is better for a team that wants a $15 per million UTF-8 bytes list price and 83 languages, and MiniMax is better for a team that wants mu-law output and a WebSocket that takes text a character at a time. Fish Audio wasn't in Design Arena's 4 August 2026 results, and MiniMax entered an older model.
Which is cheaper, Fish Audio or MiniMax?
Fish Audio is cheaper on list price, at $15 per million UTF-8 bytes on s2.1-pro against $60 to $100 per million characters on MiniMax's speech-2.8 models. Bland Speech sits between them at $40 per million characters.
Can Fish Audio and MiniMax clone a voice?
Fish Audio and MiniMax both clone a voice from a short clip. Fish Audio's instant clone needs at least 10 seconds per clip, and its professional clone needs 10 minutes or more and a confirmation of permission. MiniMax's rapid clone takes 10 seconds to 5 minutes at $1.50 a voice and deletes the voice after 7 days without use. Bland Speech's instant clone needs one clean sample of about 10 seconds.
Do Fish Audio and MiniMax return phone-ready audio?
MiniMax returns mu-law at 8 kHz as pcmu_raw and pcmu_wav. Fish Audio's documented output formats are mp3, wav, pcm, and opus, so a phone line needs a conversion step. Bland Speech returns mu-law at 8 kHz or PCM at 8 to 48 kHz.
Which text to speech API suits a voice agent on a phone line?
For a phone agent the deciding facts are mu-law output, interruption, and how the voice sounds, and Bland Speech returns mu-law at 8 kHz, interrupts a turn when a caller talks over it, and Bland Speech v3 ranked first of 13 AI voices on 4 August 2026. Fish Audio and MiniMax both sell expressive speech and cloning, and a team that bills mostly Chinese or Hindi text should check Fish Audio's byte count before comparing rates.
