Fish Audio's search traffic is mostly people looking for meme voices. If that's you, this post isn't the right one. It's for developers who called the Fish Audio API, liked the price, and are now hitting the parts that matter in production: a concurrency cap tied to how much you've prepaid, data that stays in the US by default, and an enterprise security story that's still in progress.
Below you'll find a side-by-side of the realistic alternatives, a plain comparison of Fish Audio vs ElevenLabs (the pairing most people search for), and what changes when you switch.
Fish Audio alternatives at a glance
Prices as of October 29, 2026, from each vendor's own pricing page. Feature details are from each vendor's docs on the same date. Fish Audio and AssemblyAI details are from our Fish Audio vs AssemblyAI comparison.
| Provider (model) | List price and billing unit | Languages | Concurrency | Voice cloning | 8 kHz μ-law for phone calls | Open weights | Free option |
|---|---|---|---|---|---|---|---|
| Fish Audio (S2.1 Pro) | $15 per 1M UTF-8 bytes | See vendor docs | 5 to 50 by prepaid tier, custom by contract | Yes | See vendor docs | See vendor docs | See vendor docs |
| AssemblyAI (AssemblyAI-TTS-1) | $15 per 1M characters | See vendor docs | No hard caps, no commitment | Instant, same per-character price | See vendor docs | See vendor docs | $50 in free credits |
| ElevenLabs (v4 Turbo, Flash v2.5) | $0.04 per 1K characters ($40 per 1M) | 90+ (v4 Turbo), 32 (Flash v2.5) | Concurrent requests set by plan | API; Instant from Starter, Professional from Creator | ulaw_8000 | No | Free plan, without a commercial license |
| Cartesia (Sonic 3.6) | Credit plans from $5/mo (100K credits) to $299/mo (8M); 1 credit per character | 44 | Set by plan; custom on Enterprise | Instant on paid plans; Pro clones from Startup | pcm_mulaw | No | Free plan, 20K credits/month |
| Inworld (TTS-2, TTS-2 Flash) | $25 / $15 per 1M characters (On-Demand) | 200+ | 5 (On-Demand) to 500 (Growth) concurrent requests | Instant and professional clones | Yes (MULAW) | No | On-Demand, $0 to start, pay as you go |
| Open weights (Kokoro, Chatterbox) | Your GPU and ops time | Varies by model | Whatever you provision | Varies by model | You encode it | Kokoro Apache-2.0; Chatterbox MIT | Free to run |
One row needs a footnote. Fish bills per 1M UTF-8 bytes, and AssemblyAI, ElevenLabs, Cartesia and Inworld bill per character. Accented characters in Spanish, French, German or Portuguese text count as two UTF-8 bytes. So $15 per 1M on Fish and $15 per 1M on AssemblyAI only describe the same spend for plain English text. Price a month of your own transcripts in each unit before you compare.
Who switches from Fish Audio, and why
Two groups search for Fish Audio alternatives, and they want different things.
Creators compare Fish to ElevenLabs on voice library and price for voiceovers, character voices and narration. Fish Audio serves creators as well as developers, with tools for voiceover, audiobooks and voice changing.
Developers are the second group, and their reasons are more specific:
- Concurrency is tied to prepaid spend. Fish Audio allows 5 concurrent requests until you've prepaid $100, 15 until you've prepaid $1,000, and 50 after that, with higher limits by enterprise contract. For a voice agent, every active call holds a slot while it speaks.
- Data residency is narrow. Fish Audio keeps data in the US by default, with inference in the US and Tokyo.
- Enterprise assurances are still coming. Fish Audio was founded in 2025, its SOC 2 Type II audit is in progress, and its enterprise uptime SLA is 99%.
- Some models are still in preview. Drama 3 is in preview, and so is Fish's realtime speech-to-text model.
None of this makes Fish unusable. It does mean that a team running a voice agent at scale ends up shopping, which is how most people land on this page.
Fish Audio vs ElevenLabs
This is the comparison most people type into Google, so here it is with numbers from each vendor's own pages.
| Fish Audio (S2.1 Pro) | ElevenLabs (v4 Turbo / Flash v2.5) | |
|---|---|---|
| API list price | $15 per 1M UTF-8 bytes | $0.04 per 1K characters ($40 per 1M); v4 is $0.08 per 1K |
| Billing unit | UTF-8 bytes (accented characters count as 2) | Characters |
| Languages | See vendor docs | 90+ on v4 Turbo; 32 on Flash v2.5 |
| Emotion control | Yes | Audio tags on v4 and v4 Turbo |
| Voice cloning | Yes | Instant Voice Cloning from the Starter plan; Professional from Creator |
| Concurrency | 5 / 15 / 50 by prepaid amount | Set by subscription plan |
| Phone audio | See vendor docs | ulaw_8000 output |
| Open weights | See vendor docs | None |
| Free tier | See vendor docs | Free plan with 10K credits; no commercial license until Starter |
Sources: Fish Audio vs AssemblyAI, ElevenLabs API pricing, ElevenLabs models, ElevenLabs plans. Prices as of October 29, 2026.
So how do the prices compare? On list price per million units, Fish S2.1 Pro is lower than ElevenLabs v4 Turbo and Flash v2.5. The gap narrows when your text has a lot of accented characters, because those cost Fish 2 bytes each. ElevenLabs' limits are different ones: Flash v2.5, its low-latency model, covers 32 languages, and the free plan doesn't include a commercial license.
If you're building a voice agent, neither one is your only option. The rest of this post is about the ones built for that job.
The alternatives, one by one
AssemblyAI-TTS-1: for voice agents
AssemblyAI-TTS-1 is a text-to-speech model built for production voice agents. It's trained on millions of hours of audio, including real conversational data. It runs with unlimited concurrency, with no prepaid tiers or contracts, and it costs $15 per 1M characters on the pricing page.
Against Fish, the independent Cekura text-to-speech benchmark shows, as of October 28, 2026:
- Streaming time to first audio: 159 ms for AssemblyAI-TTS-1 vs 321 ms for Fish Audio S2.1 Pro.
- Codes and addresses read correctly: 92% vs 86% for Fish Audio S2.1 Pro.
Other things that matter when you move off Fish:
- Voice cloning: instant, at the same per-character price as every other voice.
- Barge-in: streaming WebSocket and HTTP endpoints, word-level timestamps, and a cancel message that stops audio when the caller cuts in.
- Data residency: text-to-speech processing in the US or the EU at the same price, with no regional surcharge.
- Frameworks: LiveKit and Pipecat reference agents to drop it into an existing agent.
- One key for the voice stack: Universal-3.6 Pro Realtime for speech-to-text, LLM Gateway, and AssemblyAI-TTS-1 on one API key.
- Compliance: SOC 2 Type 2, ISO 27001, PCI DSS and GDPR, with 99.9% uptime.
The gap, stated flatly:
- Emotion control is coming soon on AssemblyAI-TTS-1. Fish Audio has it today.
For a head-to-head on the brand pair, see Fish Audio vs AssemblyAI.
Paste a few lines your agent actually says and compare voices before you change any code.
ElevenLabs
ElevenLabs is the default comparison for Fish, covered above. For a voice agent, the practical limits are these:
- The low-latency Flash v2.5 model covers 32 languages, and v4 Turbo lists at $0.04 per 1K characters.
- Concurrency is set by subscription plan.
- The free plan has no commercial license.
If you're weighing ElevenLabs specifically for an agent, our AssemblyAI vs ElevenLabs comparison goes model by model.
Cartesia
Cartesia sells Sonic 3.6 through credit plans, at 1 credit per character, from a free tier with 20K credits a month to $299/month for 8M credits. It covers 44 languages and outputs μ-law for telephony. Concurrency is set by plan, with custom limits only on Enterprise, so a burst of calls means planning your plan around peak traffic. See Cartesia pricing.
Inworld
Inworld lists TTS-2 at $25 per 1M characters and TTS-2 Flash at $15 per 1M on its pay-as-you-go On-Demand tier. Concurrency runs from 5 concurrent requests on On-Demand to 500 on Growth, so the plan you pay for sets how many calls can speak at once. See Inworld pricing.
Open-weight models
If self-hosting is the point, look at models with permissive licenses. Kokoro is Apache-2.0, and Chatterbox is MIT. You trade the per-character bill for GPU capacity, scaling, monitoring and model updates that you run yourself, and none of them gives you μ-law or a hosted API.
Switching from Fish Audio: what changes
Here's what changes on the settings that matter most for a voice agent:
| Area | Fish Audio | AssemblyAI-TTS-1 |
|---|---|---|
| Concurrency | 5 to 50 by prepaid tier, custom by contract | No hard caps, no commitment |
| Price | $15 per 1M UTF-8 bytes (S2.1 Pro); $50 per 1M bytes (Drama 3) | $15 per 1M characters, every voice included |
| Voice cloning | Yes | Instant, same per-character price |
| Emotion control | Yes | Coming soon |
| Data residency | US by default; inference in US and Tokyo | US or EU, same price |
| Uptime | 99% SLA with premium support | 99.9% |
The switch itself is short. Clone the LiveKit or Pipecat reference agent and point the text-to-speech stage at AssemblyAI-TTS-1. It streams audio from the first byte, supports a cancel message for barge-in, and returns word timestamps, so it drops into an existing turn loop. Then run it on a slice of your live calls next to your current voice and compare completion rates.
Get an API key, pick a voice, and run your agent's own lines through AssemblyAI-TTS-1.
Which Fish Audio alternative to pick
Here's the short version:
- You're a creator making voiceovers or character voices: ElevenLabs is the usual comparison, but read each free plan's commercial terms first.
- You want to own the model: use a permissively licensed open model (Kokoro, Chatterbox).
- You're shipping a voice agent: the concurrency model, telephony output and latency matter more than the size of the voice library.
That last group is where the list price stops being the deciding number. Two vendors can both charge $15 per 1M, and you'll still feel the difference on your busiest hour, when one has capped you at the concurrency you prepaid for and the other hasn't. Before you choose, run the same 50 agent lines through your top two picks, measure time to first audio from your own region, and count how often each one misreads a confirmation code. Our guide to evaluating TTS for voice agents has a test plan, and TTS latency explained covers what to measure. For a wider field, see our roundup of text-to-speech APIs.
Run your agent's real lines, measure time to first audio from your region, and keep the numbers.
Frequently Asked Questions (FAQs)
Fish Audio vs ElevenLabs: how does pricing compare?
Fish Audio S2.1 Pro lists at $15 per 1M UTF-8 bytes, and ElevenLabs v4 Turbo and Flash v2.5 list at $0.04 per 1K characters, which is $40 per 1M (prices as of October 29, 2026). The units differ: Fish counts accented characters as 2 bytes, so text in Spanish, French or German costs more per character on Fish than plain English does. Price your own text in both units.
What's the best Fish Audio alternative for voice agents?
For a voice agent, look for concurrency without caps, fast first audio and accurate codes. AssemblyAI-TTS-1 has no hard caps on concurrency on pay-as-you-go, and on Cekura's benchmark (October 28, 2026) it starts streaming audio in 159 ms against 321 ms for Fish Audio S2.1 Pro and reads 92% of codes and addresses correctly against 86%. It costs $15 per 1M characters.
Does Fish Audio have a concurrency limit?
Yes. Fish Audio allows 5 concurrent requests until you've prepaid $100, 15 until you've prepaid $1,000, and 50 after that, with higher limits by enterprise contract. AssemblyAI has no hard caps on concurrency on pay-as-you-go, so inbound and outbound spikes never wait on a contract.
Can I run text-to-speech in the EU instead of Fish Audio's US default?
Yes, with AssemblyAI. Text-to-speech runs in the US or the EU at the same price, and the voice library and voice cloning are available in both regions. Fish Audio keeps data in the US by default, with inference in the US and Tokyo.
How do I switch from Fish Audio to AssemblyAI?
Clone the LiveKit or Pipecat reference agent and point the text-to-speech stage at AssemblyAI-TTS-1. It streams audio from the first byte, supports a cancel message for barge-in, and returns word timestamps, so it drops into an existing turn loop. Run it on a slice of your live calls next to your current voice and compare completion rates.
