# Speak in a voice cloned from a short recording

> Speak in a voice cloned from a short recording: 2 companies fully do it, 16 with a limit, 41 carry the tag. Every listing in Text-to-speech with the reason and a source.

- Canonical: https://www.anchorterminal.com/niches/text-to-speech.clone-voice-speak/
- Markdown: https://www.anchorterminal.com/niches/text-to-speech.clone-voice-speak/index.md (~2,550 tokens)
- Slim: https://www.anchorterminal.com/niches/text-to-speech.clone-voice-speak/index.min.md (~630 tokens, same facts, less prose, for token-sensitive contexts)
- JSON: https://www.anchorterminal.com/niches/text-to-speech.clone-voice-speak/index.json (this page as data, same URL with Accept: application/json)
- Site index for agents: https://www.anchorterminal.com/llms.txt (the key docs and every listing in one file, at most 16 MB, about 4 million tokens: https://www.anchorterminal.com/llms-full.txt)
- API: https://www.anchorterminal.com/api/v1/index.json
- MCP: https://www.anchorterminal.com/mcp (the whole catalogue as tools, no key)
- DNS: dig +short TXT <slug>.grade.anchorterminal.com (the grade as a DNS TXT record, for a sandbox with no HTTP where names still resolve; also top., best.<category>. and <a>.vs.<b>. under the zone, and the same under grade.letme.dev)
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)
- Updated: 2026-11-15

Two or three companies' docs show a full fit.

| Companies, full fit | With a limit | Carry the tag | Share of the category's work | First-hand accounts |
| --- | --- | --- | --- | --- |
| 2 | 16 | 41 | 10 | 7 |

## What counts as done

A voice id is created from the clip in one call without a human review step, speaks the text in the clip's language and at least one other, and the clone keeps working when the model version changes.

Measured, for our own tests: speaker similarity a listener accepts as the same person.

How agents ask for it: make it sound like this recording; clone my voice and read these messages; use the founder's voice for the narration

## Does the job

The docs show it meets the test above, and nothing we read says otherwise.

- **Cartesia Sonic TTS API + MCP (Cartesia AI, Inc.)**: Clone Voice takes a 10 to 60 second clip and its language and returns a voice with no consent or verification step; other languages are added with Add Voice Accents, and dated snapshots keep output fixed across model updates.
- **Fish Audio TTS API (Hanabi AI Inc.)**: POST /model with train_mode fast creates a voice id from an uploaded clip in one call with no verification field, or inline references clone on the fly from a 10 to 30 second clip; the id works across s2.1-pro, s2-pro and drama-3. acc-17 found it the most natural.

## Does the job, with a limit

It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.

- **Audio8-TTS-Realistic 0.6B (Edge0 (edge0.ai), which names no legal entity or location on its site. The technical report's authors give their affiliation as AutoArk, and commit metadata in Edge0-AI/Audio8_TTS uses an autoark.ai domain)**: One clip and its exact transcript clone a voice in the processor call with no verification, in 23 languages, but similarity scores (0.66 to 0.75) are Edge0's own, and a reference is re-sent per request rather than stored as a voice id.
- **Azure AI Speech text-to-speech (Microsoft Corporation)**: Personal voice clones from a 5 to 90 second prompt and speaks 91 languages, but every clone needs a recorded consent statement and the API is restricted to customers approved through an intake form, a human step before the first clone.
- **ElevenLabs Text to Speech API + MCP (Eleven Labs Inc.)**: POST /v1/voices/ivc/create takes audio files and returns a voice_id at once, with requires_verification only when the sample matches a protected library voice, and the clone speaks 90-plus languages on v4; but the listing's note says older clones need retraining for v4, so a clone is tied to a model generation.
- **Fish Audio Drama 3 (39 AI, Inc. (the licensor named in the Fish Audio Community License). fish.audio's terms of use name Hanabi AI Inc. as the provider of the hosted services)**: A reference clip and transcript in the request clone a voice with no verification step and the card claims lower cross-lingual error rates, but acc-31 measured 6 to 20 percent failed takes with cloned French voices and acc-32 saw speaker 1 rendered in speaker 0's voice.
- **Fish Audio S2 Pro (39 AI, Inc. (the licensor named in the Fish Audio Research License). fish.audio's terms of use name Hanabi AI Inc. as the provider of the hosted services)**: POST /v1/references/add saves a voice from a clip and transcript with no verification step and the card says 10 to 30 seconds suffice across 80-plus languages, but the Research License bars commercial use and no measured similarity is independent.
- **Google Cloud Text-to-Speech (Google LLC)**: Chirp 3 Instant Custom Voice clones from a short recording plus a spoken consent statement and transfers an en-US key to six other locales, at $60 per million characters, but access is restricted to allow-listed users who must contact sales first.
- **HeyGen Voice (HeyGen Technology, Inc.)**: POST /v3/models/audio/voices with mode instant returns a voice_id usually ACTIVE within seconds, free during the preview, 20 calls a day on free workspaces, in 41 languages; but instant voices made before HeyGen Voice 1 had to be cloned again, so a model change can break a clone.
- **Inworld Realtime TTS-2 (Theai, Inc. dba Inworld AI)**: Instant cloning makes a voice from 3 to 30 seconds through the API with no review step, and TTS-2 speaks 200-plus languages, but each plan caps custom voices (creation fails with 400 at the limit) and inworld-tts-2 is updated in place with no dated snapshot to pin.
- **Ming-3.0-tts-0.5B (inclusionAI (the LICENSE's copyright holder; source headers read Copyright (c) Ant Group))**: The TTS task clones from prompt_wav and its transcript in one call with no review step, in Chinese or English plus five languages via MULTILINGUAL, but the card calls French speaker similarity the clearest gap and a discussion reports noise on Apple Silicon.
- **Ming-3.0-tts-tiny (inclusionAI, Ant Group (named in the site footer of inclusion-ai.org; the LICENSE copyright line reads inclusionAI))**: ref_audio plus ref_text in one request clones with no review step across seven languages, but the vLLM recipe carries a second-speaker similarity warning and the card lists speaker similarity as weaker than content accuracy.
- **MiniMax Speech (T2A) (Nanonoble Pte. Ltd.)**: Rapid cloning takes a 10-second to 5-minute clip in one call with no review step, $1.5 per voice charged on first use, language_boost for the clip's language; but a clone not used within 7 days is deleted, and paid preview text is the only check before that.
- **Murf TTS API + MCP (Murf Inc.)**: Instant cloning from one clip of up to 30 seconds returns a requestId to poll to COMPLETED, speaks any Falcon 2.1 locale at the standard rate, with no review step; but it is Enterprise plan only and the endpoints return 403 until enabled per workspace.
- **OmniVoice (k2-fsa, the Next-gen Kaldi open-source project. The paper's authors are at Xiaomi Corp.; no company is named as the licensor)**: A 3 to 10 second clip clones a voice in one generate() call, with Whisper transcribing if ref_text is omitted and a prompt saved for reuse, in 600-plus languages; but cross-lingual output carries the reference language's accent and the weights are non-commercial.
- **Smallest.ai Lightning TTS (Smallest, Inc.)**: One POST /waves/v1/voice-cloning with a 5 to 15 second clip returns a voiceId with no review step and cross-lingual output is supported, but clones made on lightning-v3.1 stop working on 30 November 2026 and must be recreated on Pro.
- **Soniox Text-to-Speech (Soniox Inc.)**: POST /v1/voices takes a name and a clip, processes it in seconds with no review step and the clone speaks all 60-plus languages, but a new model leaves existing voices not_computed until the agent calls recompute, and an organisation holds 20 voices.
- **VieNeu-TTS v3 Turbo (Phạm Nguyễn Ngọc Bảo, an individual (named on the card as author; Hugging Face account pnnbao-ump, GitHub account pnnbao97). No company found)**: A 3 to 8 second clip clones a voice in one infer() call or POST /v1/voices with no verification, but the server keeps clones in process memory, only Vietnamese and English are spoken, and the card warns the preset roster changes between SDK versions.

## Couldn't check

The docs were closed to us, or silent. These carry no mark either way.

- **Resemble AI Text-to-Speech API (Resemble AI, Inc.)**: The voice-generation docs index returned no voice creation page and the OpenAPI file covers Detect endpoints only, so no create-voice endpoint, clip length or review step could be read.

- not this job: Amazon Polly (Amazon Web Services, Inc.): No voice creation or cloning operation in the API; the docs index lists only lexicon, voice-list and synthesis operations.
- not this job: Deepgram Text-to-Speech (Aura-2, Flux TTS) (Deepgram, Inc.): No voice creation endpoint; voices are fixed catalogue models named flux-{voice}-{language}.
- not this job: Qwen3-TTS 12Hz 1.7B CustomVoice (Alibaba Cloud (the copyright holder in the Qwen licences; RDAP names Alibaba Cloud (Singapore) Private Limited as registrant of qwen.ai, from the first batch's check)): CustomVoice has nine preset speakers; cloning from a clip is the Base edition, a different checkpoint.
- not this job: Rime TTS API + MCP (Rime Labs, Inc.): Cloning is an Enterprise or Growth service: 30 to 60 minutes of recordings delivered to Rime's team and trained offline; no API creates a voice from a clip.

## What stops agents finishing

- a consent or identity verification step a human must complete before a clone can be used
- a purchased slot per professional voice (HeyGen) or cloning only on Enterprise (Rime)
- clones tied to a model that is retiring (Smallest Lightning v3.1 clones stop on 30 November 2026; ElevenLabs clones need retraining for v4)
- runaway or silent generations with cloned voices in some languages (acc-31: 6 to 20 percent in French on Drama 3)

## Around it

Before: speech-to-text.long-recording-batch, storage.read-file. After: text-to-speech.long-form-narration, text-to-speech.voice-agent-turn.

## Where this comes from

Weight 10 of 100 in Text-to-speech: 7 of 46 accounts tagged (acc-17, 22, 23, 31, 32, 41, 43); cloning was the reason three people picked their tool and the thing that failed in the two GitHub issues

As data: https://www.anchorterminal.com/api/v1/use-cases/text-to-speech.clone-voice-speak.json, or the MCP tool get_use_case with id text-to-speech.clone-voice-speak. Draft, read 2026-11-12 from public docs and first-hand accounts, not yet reviewed by a person.

---

Query this catalogue directly over MCP: https://www.anchorterminal.com/mcp (Streamable HTTP, no key). How: https://www.anchorterminal.com/agents/#mcp
