# Narrate a long text (article, book, podcast script) in one steady voice

> Narrate a long text (article, book, podcast script) in one steady voice: 3 companies fully do it, 20 with a limit, 40 carry the tag. Every listing in Text-to-speech with the reason and a source.

- Canonical: https://www.anchorterminal.com/niches/text-to-speech.long-form-narration/
- Markdown: https://www.anchorterminal.com/niches/text-to-speech.long-form-narration/index.md (~2,650 tokens)
- Slim: https://www.anchorterminal.com/niches/text-to-speech.long-form-narration/index.min.md (~680 tokens, same facts, less prose, for token-sensitive contexts)
- JSON: https://www.anchorterminal.com/niches/text-to-speech.long-form-narration/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 |
| --- | --- | --- | --- | --- |
| 3 | 20 | 40 | 18 | 14 |

## What counts as done

One audio file per chapter, headings paused, and the whole book priced in advance from the published rate.

Measured, for our own tests: the voice, pace and loudness match between the first and last minute (no voice switch, no change of speaker on a blind listen), and no sentence is missing or duplicated against the text.

How agents ask for it: read this article to me; make an audiobook of this EPUB; turn this script into a podcast episode

## Does the job

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

- **Amazon Polly (Amazon Web Services, Inc.)**: StartSpeechSynthesisTask takes up to 100,000 billed characters and writes to S3; the published rate is $4 per million standard, $16 neural, $100 long-form; AWS warns model updates can shift generative and long-form voices over a long project.
- **Azure AI Speech text-to-speech (Microsoft Corporation)**: Real-time requests stop at 10 minutes; the GA batch synthesis API takes plain text or SSML (lexicons included) for audio longer than that and is billed per character on the same meters.
- **MiniMax Speech (T2A) (Nanonoble Pte. Ltd.)**: The async API takes 100,000 characters per text and 1,000,000 per task at the same $100 (HD) or $60 (Turbo) per million characters as sync, with Chinese counted twice; acc-12 narrates with Speech-02-HD and rates it far above Chatterbox and Kokoro, though not cheap.

## 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)**: Free to run and the voice is fixed by one reference, but reference, transcript and text share 2,048 positions, so a chapter must be split into short utterances with no context or previous-text mechanism between them.
- **Cartesia Sonic TTS API + MCP (Cartesia AI, Inc.)**: No long-audio or async route; contexts hold prosody across inputs on one connection and a dated snapshot fixes the model, but acc-01 found emotional delivery drifting across 50-plus scenes, and the per-plan credit prices sit on a page that redirected.
- **Deepgram Text-to-Speech (Aura-2, Flux TTS) (Deepgram, Inc.)**: Batch POST /v2/speak renders one block of text to a file and cross-turn context keeps prosody on the WebSocket, but there is no async or long-audio route, no seed, and the batch character limit is not stated on the pages read (Aura-2 stops at 2,000).
- **ElevenLabs Text to Speech API + MCP (Eleven Labs Inc.)**: previous_text, next_text, previous_request_ids and a best-effort seed keep chunks continuous and Multilingual v2 is named most stable for long form, but requests stop at 10,000 characters (40,000 on Flash) with no async route, and acc-46 and acc-02 price a book at USD 75 or a couple of hundred euros.
- **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)**: Long text is split into chunks of about 300 bytes with the first and latest chunk fed back as context, which is the consistency mechanism the use case wants, but acc-31 reports non-deterministic runaway takes and acc-33 Chinese generation stopping at the first punctuation mark.
- **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)**: acc-20 and acc-21 recommend it for whole-ebook runs on prosody and long text is chunked by chunk_length 200 at no per-word cost, but the licence is research and non-commercial only and --seed does not fix the timbre between chunks.
- **Fish Audio TTS API (Hanabi AI Inc.)**: Price is computable per UTF-8 byte and chunk_length with condition_on_previous_chunks carries context between chunks, but the OpenAPI file sets no text limit and there is no async or long-audio route, so chapter-length requests are untested territory.
- **Google Cloud Text-to-Speech (Google LLC)**: synthesizeLongAudio takes up to 1 million bytes to Cloud Storage and sync requests stop at 5,000 bytes, with Chirp 3 HD at $30 and Neural2 at $16 per million characters; but Gemini-TTS has no seed and acc-16 heard voices vary between generations, while acc-14 and acc-46 use it cheaply.
- **HeyGen Voice (HeyGen Technology, Inc.)**: text is capped at 10,000 characters a request with no async or long-audio route, so chapters must be chunked and stitched; no previous-text or context mechanism, seed only on professional voices, and the per-character rate is not on the pages read today.
- **Inworld Realtime TTS-2 (Theai, Inc. dba Inworld AI)**: synthesizeAsync takes one text up to 100,000 characters as a job, priced at $25 per million characters on On-Demand, but On-Demand jobs stop at 10,000 characters and no previous-text or seed mechanism is documented for keeping the voice steady across chapters.
- **Ming-3.0-tts-0.5B (inclusionAI (the LICENSE's copyright holder; source headers read Copyright (c) Ant Group))**: Runs locally at no per-character cost with built-in voices, and generate() defaults to temperature 0 and seed 1234, but one task per call with no documented text limit or context mechanism, so the agent chunks and checks voice drift itself.
- **Ming-3.0-tts-tiny (inclusionAI, Ant Group (named in the site footer of inclusion-ai.org; the LICENSE copyright line reads inclusionAI))**: Free to run with a seed on the endpoint and built-in voices, but no text limit, context or previous-text mechanism is documented, so the agent chunks and checks for drift; the card notes stability with several tags in one sentence is still improving.
- **Murf TTS API + MCP (Murf Inc.)**: Gen2 synthesize returns a file link that lasts 72 hours and Falcon 2.1 streams at 1 cent per 1,000 characters, but no async or long-audio route exists, no per-request cap or consistency mechanism is documented, and the agent must chunk and stitch.
- **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)**: RTF as low as 0.025 is claimed and a saved VoiceClonePrompt keeps the same reference across calls, but there is no long-text chunking or previous-text context in generate(), and the weights are CC-BY-NC, so only a personal audiobook qualifies.
- **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))**: Free with nine fixed voices and batch lists in one call, but no context or previous-text mechanism, max_new_tokens defaults to 2048, and acc-19 found it slow and occasionally switching to a different voice mid-text.
- **Resemble AI Text-to-Speech API (Resemble AI, Inc.)**: Synchronous and HTTP streaming take 2,000 characters a request and the WebSocket 3,000, with no async or long-audio route, so chapters are chunked and stitched; the seed is returned but a steady voice across chunks is not documented, and the price could not be read.
- **Rime TTS API + MCP (Rime Labs, Inc.)**: Billing is per minute of audio at $0.03 (Mist v3) or $0.05 (Coda) with about 1,000 characters a minute, so a book can be estimated; but there is no long-audio route, no context or seed to hold the voice between chunks, and Coda drops timestamps past 180 seconds.
- **Smallest.ai Lightning TTS (Smallest, Inc.)**: Both model cards cap a request at 250 characters with no async or long-audio route, so the agent chunks and stitches; continuations carry prosody across fragments on the WebSocket only. Price is computable at $0.175 per 10,000 characters.
- **Soniox Text-to-Speech (Soniox Inc.)**: Generated audio is capped at 2 minutes per request or stream, fixed and with no async route, so the agent chunks and stitches; no previous-text or seed mechanism is documented. Price is per token, about $0.70 an hour.
- **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)**: Up to 20,000 characters per request split at sentence boundaries with the voice fixed by a preset embedding, free under Apache 2.0, but only Vietnamese and English, and there is no previous-text or seed mechanism between chunks.

## What stops agents finishing

- a per-request cap (2 minutes at Soniox, 2,000 characters at Deepgram and Resemble) with no async or long-audio route, so the agent must chunk and stitch
- voice drifting between chunks or generations, which a seed or previous-text context does not fully fix
- a book-length price far above the alternatives (acc-46 puts 250k words at about USD 75 on ElevenLabs against USD 15 on Gemini TTS)
- a free tier that stops after minutes a day (acc-16: about 90 minutes a day)

## Around it

Before: web-scrapers.page-to-markdown, document-extraction.pdf-to-text. After: storage.upload-file, podcast.publish-feed.

## Where this comes from

Weight 18 of 100 in Text-to-speech: 14 of 46 accounts tagged (acc-02, 10, 12, 14, 15, 16, 19, 20, 21, 22, 25, 39, 45, 46); the most common hobbyist and accessibility job on Hacker News, where price per book decided the tool

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