# Run text-to-speech on the agent's own hardware for cost or privacy

> Run text-to-speech on the agent's own hardware for cost or privacy: 1 companies fully do it, 7 with a limit, 127 carry the tag. Every listing in Text-to-speech with the reason and a source.

- Canonical: https://www.anchorterminal.com/niches/text-to-speech.self-hosted-local/
- Markdown: https://www.anchorterminal.com/niches/text-to-speech.self-hosted-local/index.md (~2,000 tokens)
- Slim: https://www.anchorterminal.com/niches/text-to-speech.self-hosted-local/index.min.md (~530 tokens, same facts, less prose, for token-sensitive contexts)
- JSON: https://www.anchorterminal.com/niches/text-to-speech.self-hosted-local/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

One company's docs show a full fit.

| Companies, full fit | With a limit | Carry the tag | Share of the category's work | First-hand accounts |
| --- | --- | --- | --- | --- |
| 1 | 7 | 127 | 12 | 15 |

## What counts as done

A model installs from its documented command, serves an OpenAI-style /v1/audio/speech or equivalent endpoint, and the licence allows the stated use.

Measured, for our own tests: it synthesises a 200-word passage faster than real time on the stated hardware.

How agents ask for it: speak this without sending it to a cloud API; set up a local TTS server I can call like OpenAI; which open model can I run on my 24 GB card

## Does the job

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

- **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)**: pip install vieneu then python -m apps.openai_speech serves /v1/audio/speech; RTF 0.49 on an RTX 3060 using 1.1 GB of VRAM and 0.55 on a 6-core CPU through ONNX, with Apache 2.0 covering commercial use and the preset voices.

## 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)**: Loads with transformers 4.57 and trust_remote_code under Apache 2.0 and the adapter serves /v1/audio/speech, but the card gives no GPU memory figure or real-time factor for Realistic Air (tested on an H20) and the adapter needs a pinned SGLang Omni commit.
- **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)**: Installs and serves /v1/audio/speech through SGLang-Omni or vLLM-Omni from source, with RTF 0.131 at one stream on an H200; but it needs an NVIDIA GPU with at least 48 GB for FP8, outside the 8 to 32 GB input, and commercial use is free only under USD 500,000 revenue.
- **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)**: Installs with uv sync and serves POST /v1/tts (msgpack, not OpenAI-shaped) on an NVIDIA GPU with 24 GB, CPU, ROCm or Intel XPU, with a vLLM-Omni recipe linked from the README; but the weights are research and non-commercial only.
- **Ming-3.0-tts-0.5B (inclusionAI (the LICENSE's copyright holder; source headers read Copyright (c) Ant Group))**: git clone and pip install -r requirements.txt are documented and the licence is MIT, but there is no OpenAI-style endpoint for this checkpoint (the vLLM-Omni recipe covers tiny only), and the publisher tested only H800 and H20 GPUs with no real-time figure.
- **Ming-3.0-tts-tiny (inclusionAI, Ant Group (named in the site footer of inclusion-ai.org; the LICENSE copyright line reads inclusionAI))**: vllm serve --omni and sgl-omni serve both give an OpenAI-compatible /v1/audio/speech under MIT, but the recipe names an H100 or H200, the weights are 17.7 GB BF16 with no quantised build, and no faster-than-real-time figure is given.
- **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)**: pip install omnivoice runs on CUDA, Apple MPS or Intel XPU (tested on a 4 GB Arc A310), but there is no OpenAI-style endpoint, only omnivoice-demo and CLI tools, and the pre-trained weights are CC-BY-NC, so commercial use is not granted.
- **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))**: pip install -U qwen-tts is documented under Apache 2.0 and vLLM-Omni's vllm serve gives an OpenAI endpoint, but the README still says only offline vLLM inference exists, no GPU memory is stated, and acc-19 reports slow generation.

- not this job: Amazon Polly (Amazon Web Services, Inc.): A hosted AWS API; no weights or local runtime.
- not this job: Azure AI Speech text-to-speech (Microsoft Corporation): A hosted Azure service; the pages read document no open weights or local endpoint, and the embedded-speech meter in the price list is a licensed container, not a model to install.
- not this job: Cartesia Sonic TTS API + MCP (Cartesia AI, Inc.): Hosted API only; no weights.
- not this job: Deepgram Text-to-Speech (Aura-2, Flux TTS) (Deepgram, Inc.): Hosted API; no open weights or local install in the docs.
- not this job: ElevenLabs Text to Speech API + MCP (Eleven Labs Inc.): Hosted API only; no weights.
- not this job: Fish Audio TTS API (Hanabi AI Inc.): A hosted API; the open weights are the separate Drama 3 and S2 Pro listings.
- not this job: Google Cloud Text-to-Speech (Google LLC): Hosted API; no weights.
- not this job: HeyGen Voice (HeyGen Technology, Inc.): Hosted API only; no weights or local runtime.
- not this job: Inworld Realtime TTS-2 (Theai, Inc. dba Inworld AI): Hosted API only; no weights or local runtime.
- not this job: MiniMax Speech (T2A) (Nanonoble Pte. Ltd.): Hosted API only; the self-host guide covers Music 3, not speech.
- not this job: Murf TTS API + MCP (Murf Inc.): On-premise Falcon 2.1 is a licensed Enterprise server, not open weights.
- not this job: Resemble AI Text-to-Speech API (Resemble AI, Inc.): Hosted API only; no weights or local runtime.
- not this job: Rime TTS API + MCP (Rime Labs, Inc.): On-prem is a licensed Docker image needing a runtime licence from Rime and an NVIDIA T4 or better, not open weights.
- not this job: Smallest.ai Lightning TTS (Smallest, Inc.): Hosted API only; on-premises deployment is an Enterprise line on the pricing page, not weights an agent can run.
- not this job: Soniox Text-to-Speech (Soniox Inc.): Hosted API only; the docs offer regional hosts, not weights or an on-premises build.

## What stops agents finishing

- NVIDIA only with 24 to 32 GB needed (Drama 3, Ming tiny, S2 Pro) and no quantised build from the publisher
- no released mainstream runtime, only the publisher's scripts with trust_remote_code (Ming, Audio8, OmniVoice)
- non-commercial or revenue-capped licence (S2 Pro, OmniVoice weights, Drama 3 above USD 500,000)
- no streaming in the reference package (Qwen3-TTS, Ming 0.5B), so sentence-level chunking has to be built by the agent

## Around it

Before: inference.local-serve. After: text-to-speech.voice-agent-turn, text-to-speech.long-form-narration.

## Where this comes from

Weight 12 of 100 in Text-to-speech: 15 of 46 accounts tagged (acc-03, 07, 10, 11, 19, 21, 22, 24, 25, 31, 32, 33, 42, 43, 44), the largest raw count; Hacker News over-counts local users, so this is the least certain weight and is set below its raw share

As data: https://www.anchorterminal.com/api/v1/use-cases/text-to-speech.self-hosted-local.json, or the MCP tool get_use_case with id text-to-speech.self-hosted-local. 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
