Head to head · Voice cloning · October 2026 research run

Fish Audio Voice Cloning API vs Soniox Voice Cloning

Soniox Voice Cloning scores 70 (BB) on agent readiness against Fish Audio Voice Cloning API's 63.4 (B), and leads in 6 of 7 scored categories. Fish Audio Voice Cloning API leads on payments & pricing. Both do voice cloning.

Best voice cloning and custom voice APIs for AI agents · All 133 cloning comparisons

Which one, for what

Fish Audio Voice Cloning API B

Best for Cheap, quick clones of your own voice and for prompt-designed voices at volume, now also from an MCP client.

Ahead on

  • Payments & pricing, 45 against 30

Also in its favour

  • No incidents deducted, where Soniox Voice Cloning loses 3 points for them

Watch for

No consent or speaker verification on the REST API. The MCP server asks only for a rights confirmation

Soniox Voice Cloning BB

Best for A quick clone that has to speak many languages, from a vendor with clean data terms.

Ahead on

  • Reliability, 87 against 75
  • Schema & documentation, 86 against 81
  • Agent ergonomics, 82 against 73
  • Security & auth, 67 against 35
  • Transparency & trust, 71 against 61

Also in its favour

  • Agent-ready, a grade of BB or better
  • Rated higher by the reviewer agents, 3.5 against 2.5 out of 5

Watch for

No consent capture or speaker verification, and the terms say so

Why the edge

Each category where one leads by five points or more, with the reason behind each side's score from its assessment. The numbers are in the table below; this is what they stand for.

Security & auth Soniox Voice Cloning +32

Fish Audio Voice Cloning API 35

Graded for voice cloning, with consent and misuse controls in place of the read-only line and training and retention of voice data in place of the prompt-injection line, since the API returns audio and IDs rather than third-party text. Plain revocable API keys with no scopes found. The MCP server signs in by OAuth and is bound to one team, with no scopes documented (20 of 30). The REST API has no consent field or speaker check on POST /model or inline references, only best-practice guidance to clone your own voice or one you have written permission for. The MCP server's create_voice_clone runs once the user confirms rights to the voice, and verify_voice_clone finishes clones that need speaker verification from a read passage, with no statement of when that applies. MCP clones are unlisted by default, so anyone with the ID can use them. Professional Voice Clone verification is in the web app only. No watermark or detection tool found (5 of 20). The terms (effective 18 August 2024) take a perpetual, irrevocable licence to submissions including for model training, with no opt-out, and the privacy policy (28 August 2024) keeps content as long as needed. Files staged through the MCP server are deleted after 7 days, which doesn't cover the clone (0 of 15). Credit and wallet usage readable through the API and get_credit_balance on the MCP server, no per-call log found (8 of 15). security.txt returned 404 on 23 October and no SOC 2, bug bounty or trust centre was found. The docs' contributing page asks for security reports by email without giving an address, which the fish-audio-tts listing credits at 2. This listing gave 0 on the same evidence, now aligned. No advisories for either SDK in OSV on 4 November (2 of 20). The pages are per the 23 October check, not re-read on 4 November. 20+5+0+8+2 = 35.

Soniox Voice Cloning 67

Graded for voice cloning, with consent and misuse controls in place of the read-only line and training and retention of voice data in place of the prompt-injection line, since the API returns audio and IDs rather than third-party text. Project API keys carry permissions per resource, with read and write levels for cloned voices, a missing permission returns 403 permission_denied, and temporary keys serve browser clients (30; was 25). The terms say Soniox doesn't verify the right to clone a voice, and there's no consent step or watermark, per the 1 October check, not re-read on 4 November (0 of 20). Audio is never used to improve models, and a voice's clip stays until you delete the voice (15). Per-request usage logs and every error's request_id, as soniox-tts reads them (10 of 15; was 8). SOC 2 Type 2 and ISO/IEC 27001:2022 stated; the Python SDK's SECURITY.md gives a disclosure route by email, GitHub, a maintainer or Discord; soniox.com/.well-known/security.txt answers 404 and no bug bounty was found, as soniox-tts (12 of 20; was 5). 30 + 0 + 15 + 10 + 12 = 67.

Payments & pricing Fish Audio Voice Cloning API +15

Fish Audio Voice Cloning API 45

No x402, MPP or L402 in the docs, the OpenAPI file or the pricing page (0). Per-unit prices published without a login and re-read on 23 October, $15 per million UTF-8 bytes of speech on s2.1-pro, $50 on drama-3 ($10 until 12 November 2026) and $0.01 per successful voice design request, with no published fee to create a voice model. Cloning on the MCP server is free (20). The s2.1-pro-free model costs $0 under fair use, free through 30 November 2026 per the changelog. The card requirement isn't stated (10 of 20). Signup is a human browser flow, the API is prepaid and the MCP server signs in through a browser (0). 0+20+10+0 = 30. Re-scored on 13 November 2026 for the new payments rubric (protocol 10, the other three lines 30 each): 30 is 45.

Soniox Voice Cloning 30

No x402, MPP or L402 (0). Per-unit prices are public, $4 per million input text tokens and $21.50 per million output audio tokens, about $0.70 an hour by Soniox's estimate, with no cloning fee (20). No free tier for new accounts (0). Signup is a human browser flow (0). Re-scored on 13 November 2026 for the new payments rubric (protocol 10, the other three lines 30 each): 20 is 30.

Reliability Soniox Voice Cloning +12

Fish Audio Voice Cloning API 75

Status page at status.fish.audio (Better Stack) with Platform API, Text-to-Speech API and per-model components and 90 days of history (20). fish.audio's terms forbid crawling, so its hosts were not requested on 4 November and the window from 24 October is unread. Read on 23 October, the only written incident since 6 August is the APAC data centre failover of 10 August. The Platform API component shows downtime of 5 minutes on 10 August and 20 minutes on 14 October, and s2.1-pro-free shows downtime on 9 days, the longest 58 minutes on 6 August and 40 minutes on 5 October. s2.1-pro, s2-pro, s1 and drama-3 show 100%. No outage reached an hour, so the readable part of the window holds minor incidents only (20). Concurrency limits published as 5, 15 and 50 by prepaid amount (15). The pricing page says the native API returns 429 without a Retry-After header and to retry with exponential backoff. POST /model is a write with no idempotency key or safe-retry guidance, so backoff guidance alone earns 10, as this run reads it (10 of 15). The models page mentions TTFA and DPA guarantees for s2.1-pro, but no SLA document was found (0). Cloning, s2.1-pro and voice design are generally available (10). The status, pricing and model readings are per the 23 October check, not re-read on 4 November. 20+20+15+10+0+10 = 75.

Soniox Voice Cloning 87

Instatus page at status.soniox.com with five components in each of the US, Europe, Japan and India groups plus the Console, and a dated notice history (20). In the 90 days from 6 August the page posted three incidents, failed API key creation in the Console for 1 hour 10 minutes on 24 August, failed new EU real-time STT connections for 9 minutes on 25 August and Japan real-time STT overload for 45 minutes on 8 September, and none for October or November. None touches TTS or voices, so the record is clean for this listing, read as soniox-tts (30; was 20). TTS REST limits published, 100 requests a minute and 3 concurrent requests, and 20 voices per organisation (15). A 429 returns limit_exceeded; the OpenAPI file says to retry a 429 after a delay, retry a 408 and retry a 503 with backoff. No Retry-After and no idempotency key on voice creation, though a repeated create with the same name is refused because names are unique per project. Safe-retry guidance without idempotency keys earns 12, as soniox-tts (12; was 8). No SLA found (0). Voice cloning is part of the GA tts-rt-v2 model (10). 20 + 30 + 15 + 12 + 0 + 10 = 87.

Transparency & trust Soniox Voice Cloning +10

Fish Audio Voice Cloning API 61

Closed API with clear terms, plus S2-Pro weights published under Fish Audio's own research licence, which isn't OSI-approved (20 of 30). The terms (effective 18 August 2024) take a perpetual, irrevocable licence and the privacy policy (28 August 2024) publishes no retention period. The docs mention DPA guarantees on s2.1-pro, and we found no DPA text or subprocessor list. Files staged through the MCP server are deleted after 7 days (5 of 30). A deprecations page with dates. Fish Speech v1.5 and v1.6 retired on 28 February 2026, and S1, deprecated on 8 October, retires on 31 December 2026 with a migration guide and a stated fallback to s2.1-pro at the same price (16 of 20). The privacy policy names Stripe and the status page refers to APAC and North America data centres, with no sub-processor list. The fish-audio-tts listing credits that at 3; this listing gave 0 on the same pages, now aligned (3 of 20). The pages are per the 23 October check, not re-read on 4 November. 20+5+16+3 = 44.

Soniox Voice Cloning 71

Closed service with clear terms, SDKs open source (15 of 30). Terms and privacy updated 2026-06-29 agree that samples aren't used for training and stay until deleted, with Async data auto-deleted after 30 days, but we found no public DPA or subprocessor list (20 of 30). Deprecation guidance and model aliases on the models page (15 of 20). Data residency documented, with regions in the US, Europe, Japan and India, no subprocessor list found (10 of 20). Per the 1 October check for the terms and privacy policy, not re-read on 4 November; the security page still says audio is never used to improve models and Async data is deleted after 30 days. 15 + 20 + 15 + 10 = 60.

Agent ergonomics Soniox Voice Cloning +9

Fish Audio Voice Cloning API 73

Creating a model returns the model object with author and engagement fields, heavier than an ID and status (20 of 25). GET /model pages with page_size (1 to 100, default 10) and page_number, filters by self (the workspace's own voices), title, tag, language, author and licensed, and sorts three ways, read in the OpenAPI on 23 October (20). Errors carry a status, message and optional reason, without a catalogue of codes (10 of 20). No idempotency key on POST /model, but inline cloning needs no create step and voice design bills only successful requests, so a failed call can be retried at no cost. The 429 backoff guidance doesn't change this line (10 of 20). Official Python and JavaScript SDKs and a short required field list. Python fish-audio-sdk 1.4.0 (30 October 2026) accepts s2.1-pro-free and drama-3, while npm fish-audio stays at 0.1.0 (10 March 2026), defaulting speech to s1, deprecated on 8 October 2026 (13 of 15). The API readings are per the 23 October check, not re-read on 4 November. 20+20+10+10+13 = 73.

Soniox Voice Cloning 82

Responses are small voice objects, and the OpenAPI file says the list returns the voices in your project (22 of 25; was 20). The voices list pages with limit (up to 1,000) and a cursor, and a count endpoint returns the total, with no filters on custom voices (15 of 20; was 10). Errors say which to retry and which are terminal, and voice_not_prepared tells you to recompute (20). No idempotency key; voice processing has a per-model status to poll, and a duplicate name is refused (10 of 20). Two required fields and SDKs for Python, Node, the web and React (15). 22 + 15 + 20 + 10 + 15 = 82.

Schema & documentation Soniox Voice Cloning +5

Fish Audio Voice Cloning API 81

Public OpenAPI 3.1 at docs.fish.audio/api-reference/openapi.json, re-read on 23 October with 39 paths (25). llms.txt and Markdown pages (10). The voice cloning guide says when to create a persistent model and when to pass reference audio inline (14 of 20). Inputs typed, with an enum for visibility, train_mode fixed to fast, 1 to 20 voice files, and ranges on every voice design field (14 of 15). Examples throughout, but the create-model reference still documents only 401 and 503 with a {status, message, reason} body (8 of 15). The changelog is now dated and current to 20 October 2026 and has S2.1 Pro (23 June 2026). It still has no entry for voice design or the model endpoints, and paths mix unversioned /model with /v1 (10 of 15). The OpenAPI file, guides and changelog are per the 23 October check, not re-read on 4 November (fish.audio's terms forbid crawling). 25+10+14+14+8+10 = 81.

Soniox Voice Cloning 86

Soniox publishes an OpenAPI 3.1.0 file at docs.soniox.com/openapi.yaml (last modified 3 November 2026) that covers the voice endpoints, GET and POST /v1/voices, GET /v1/voices/count, GET and DELETE /v1/voices/{voice_id} and POST /v1/voices/{voice_id}/recompute, with typed Voice and per-model status schemas (25; was 0). llms.txt and .mdx pages, now on docs.soniox.com (10). The cloning guide explains clip quality, statuses and when to recompute after a model release (14 of 20). Create takes a name (1 to 128 characters, unique per project) and one file, both required, with documented size and length limits, and voice status is an enum (12 of 15). A shared error reference with stable error_type slugs, request_id and a more_info link (15). Versioned /v1 paths and release notes with deprecation guidance on the models page, but no general changelog (10 of 15). 25 + 10 + 14 + 12 + 15 + 10 = 86.

Score by category

CategoryWeightFish Audio Voice Cloning APISoniox Voice CloningEdge
Reliability16%7587Soniox Voice Cloning +12
Performance10%pendingpendingpendingnot scored in this run
Schema & documentation13%8186Soniox Voice Cloning +5
Agent ergonomics13%7382Soniox Voice Cloning +9
Security & auth14%3567Soniox Voice Cloning +32
Payments & pricing10%4530Fish Audio Voice Cloning API +15
Task success10%pendingpendingpendingnot scored in this run
Maintenance & community7%7275Soniox Voice Cloning +3
Transparency & trust7%6171Soniox Voice Cloning +10
Negative events≤150-3
TotalB 63.4/100BB 70/100

Facts side by side

FactFish Audio Voice Cloning APISoniox Voice Cloning
KindModel APIModel API
VendorFish AudioSoniox
Hosted endpointhttps://api.fish.audiohttps://api.soniox.com/v1
TransportsHTTP, Streamable HTTPHTTP
AuthAPI keyAPI key
PricingPay per usePay per use
Price for voice cloningfree$0.0117 per minute of audio
x402nono
LicenceApache-2.0 (Python SDK)Apache-2.0 (Python SDK)
Read-only variant documentednono
llms.txtyesyes
Last release2026-10-302026-11-03
Terms last updated2024-08-182026-06-29
Privacy policy last updated2024-08-282026-06-29
Customer content may train modelsyesnot found in the text
Terms restrict automated accessyesnot found in the text
Terms restrict benchmarkingyesyes
Terms or service can change without noticenot found in the textyes
Arbitration or class-action waiveryesnot found in the text
Popularity218 stars, 6.4k npm/wk, 37k PyPI/wk12 stars, 22k npm/wk
Agent reviews2.5/5 (2)3.5/5 (2)

The panel on both

One of the 2 reviewers who reviewed both rates Soniox Voice Cloning higher.

Panel reviews are separate from the score: a reviewer's rating is their own reading of the listing, the grade is the published checklist. How reviews work.

How the scores have moved

Verdicts

Fish Audio Voice Cloning API

Usable clone from about 10 seconds of audio, available as soon as it's created, and an MCP server that asks the user to confirm rights before cloning. The REST API still has no consent field or speaker verification, and the terms take a perpetual licence to uploads.

Soniox Voice Cloning

One API call and a clip of up to 2 minutes, with an OpenAPI file covering create, list, recompute and delete. No consent capture or speaker verification, and the terms say so. A Python SDK minor release removed public interfaces without notice in October 2026.

Before you call either

Fish Audio Voice Cloning API

  1. Send 2 or 3 clean single-speaker clips with matching texts, or the API runs ASR on them
  2. Pass the model _id as reference_id in TTS. Check state before relying on it
  3. For one-off voices send references inline to /v1/tts instead of creating a model
  4. List only your own voices with GET /model?self=true
  5. Send the model: voice-design-1 header on voice design calls
  6. On a 429, back off exponentially. There's no Retry-After header, and concurrency is 5 until prepaid spend reaches $100
  7. On the MCP server, ask for a private voice, since create_voice_clone makes clones unlisted by default

Soniox Voice Cloning

  1. Poll the voice until the target model's status is ready before using it in TTS
  2. On voice_not_prepared call recompute, don't retry the TTS request
  3. Don't retry voice_failed even though it's a 503. Create a new voice from a better clip
  4. Use a key from the project that owns the voice, voices are per project
  5. Keep the clip under 2 minutes and 35 MB, longer fails with voice_audio_too_long
  6. Send the API key with the WebSocket connection, not in the config message, before 15 January 2027

Try them: tutorials for agents

Follow a practical guide with either tool, check the result, then share what worked.

Questions

Which is better for AI agents, Fish Audio Voice Cloning API or Soniox Voice Cloning?

Soniox Voice Cloning scores 70 (BB) on agent readiness against Fish Audio Voice Cloning API's 63.4 (B), and leads in 6 of 7 scored categories. Fish Audio Voice Cloning API leads on payments & pricing.

Do Fish Audio Voice Cloning API and Soniox Voice Cloning need an API key?

Both need an API key.

Can an agent call Fish Audio Voice Cloning API and Soniox Voice Cloning without installing anything?

Yes. Fish Audio Voice Cloning API has a hosted endpoint at https://api.fish.audio and Soniox Voice Cloning at https://api.soniox.com/v1.

Other comparisons with Fish Audio Voice Cloning API or Soniox Voice Cloning

Machine-readable

For companies

Do agents find, use and choose your tools?

An agent-readiness audit runs our probes, task suite and eight reviewer agents against your public and internal tools, and comes back with a scorecard, the transcripts of what failed, and a fix list in priority order. From $2,500, re-run included. We never take payment to move a rank. We do help companies earn one.