# Direct the delivery: emotion, pace, pauses, several speakers in one scene

> Direct the delivery: emotion, pace, pauses, several speakers in one scene: 1 companies fully do it, 18 with a limit, 38 carry the tag. Every listing in Text-to-speech with the reason and a source.

- Canonical: https://www.anchorterminal.com/niches/text-to-speech.directed-performance/
- Markdown: https://www.anchorterminal.com/niches/text-to-speech.directed-performance/index.md (~2,450 tokens)
- Slim: https://www.anchorterminal.com/niches/text-to-speech.directed-performance/index.min.md (~630 tokens, same facts, less prose, for token-sensitive contexts)
- JSON: https://www.anchorterminal.com/niches/text-to-speech.directed-performance/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 | 18 | 38 | 9 | 7 |

## What counts as done

Direction (tags, markup or instructions) is accepted, each labelled line can be assigned its own voice, and a seed can be set.

Measured, for our own tests: each direction audibly changes the output on a blind listen, tags are never read aloud as words, each labelled line is spoken by its own assigned voice, and the result is the same on a repeat with the same seed.

How agents ask for it: read this angrily, then whisper the last line; voice this two-person dialogue; make the narrator sound tired

## Does the job

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

- **ElevenLabs Text to Speech API + MCP (Eleven Labs Inc.)**: Eleven v4 and v3 take bracketed audio tags such as [whispers], [laughs] and [strong French accent]; Text to Dialogue gives each turn its own voice_id; seed is best effort; SSML break tags are ignored on v4 and v3; acc-38 ships tags in production.

## 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.

- **Amazon Polly (Amazon Web Services, Inc.)**: Direction is SSML prosody (rate, pitch, volume, partial on neural and generative), breaks and a newscaster style on some neural voices; whisper and soft phonation are standard-only, there are no emotion tags, and the voice tag is unsupported so each speaker needs its own request.
- **Azure AI Speech text-to-speech (Microsoft Corporation)**: mstts:express-as styles with styledegree, several voice elements per document and a MultiTalker HD voice cover emotion and speakers, but HD voices take only a subset of SSML, personal voices ignore prosody and style tags, and there is no seed for a repeatable take.
- **Cartesia Sonic TTS API + MCP (Cartesia AI, Inc.)**: emotion, speed, volume and break tags inline plus generation_config change delivery, but a request speaks one voice, so a dialogue needs a context per speaker, and no seed is documented for a repeatable take.
- **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)**: Plain-language directions, point tags, paired tags and speaker tags 0 to 15 with one reference per speaker are documented for one request, but acc-32 reports the second speaker voiced in the first speaker's voice, acc-31 pause tags and ellipses triggering runaways, and --seed does not fix the timbre.
- **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)**: Free-form bracket tags such as [whisper] and [pause] and multi-speaker by reference order are documented, but --seed is stated not to fix the timbre, so a repeat with the same seed is not the same, and no account tests the second speaker.
- **Fish Audio TTS API (Hanabi AI Inc.)**: Bracket cues on the S2 family, plain-language directions and paired tags on drama-3, and speaker tags with one reference_id per speaker voice a scene in one request; but the request has no seed, so a repeat is not reproducible, and s1 parenthesis tags are read aloud on S2.
- **Google Cloud Text-to-Speech (Google LLC)**: Gemini-TTS takes a natural-language prompt for style, accent and emotion and voices multi-speaker dialogue (acc-45 made a two-host podcast), and Chirp 3 HD has [pause] markup; but there is no seed and acc-16 found voices vary between generations.
- **HeyGen Voice (HeyGen Technology, Inc.)**: 201 audio tags such as [excited] and [whispers] work on public and instant voices on all three models, but no speaker labels or dialogue mode exist and seed applies to professional voices only, so a two-voice scene needs two requests.
- **Inworld Realtime TTS-2 (Theai, Inc. dba Inworld AI)**: Bracketed instructions and [reset] steer inworld-tts-2 and are never spoken, and a script of 1 to 100 utterances each names its own voiceId, but Flash ignores steering, and there is no seed, only deliveryMode STABLE, so a repeat is not guaranteed identical.
- **Ming-3.0-tts-0.5B (inclusionAI (the LICENSE's copyright holder; source headers read Copyright (c) Ant Group))**: 14 inline paralinguistic tags, an EMOTION task, FREEFORM natural-language direction and speed, pitch and volume levels are documented, but one task per call so controls cannot combine, multi-tag stability is still improving, and multi-speaker podcast is tiny only.
- **Ming-3.0-tts-tiny (inclusionAI, Ant Group (named in the site footer of inclusion-ai.org; the LICENSE copyright line reads inclusionAI))**: Inline paralinguistic tags, emotion, freeform direction and a podcast task with speaker_N labels and one reference per speaker are documented, but tasks cannot combine, and the vLLM recipe warns the second speaker's similarity needs listening.
- **MiniMax Speech (T2A) (Nanonoble Pte. Ltd.)**: emotion sets one of nine moods per request, speech-2.8 reads interjection tags such as (laughs), and <#x#> inserts timed pauses, but there are no speaker labels for a two-voice scene and no seed for a repeatable take.
- **Murf TTS API + MCP (Murf Inc.)**: Gen2 takes a style such as Angry or Sad and [pause 1s] markers, but Falcon 2.1 speaks only Conversational, there are no inline emotion tags, no speaker labels and no seed.
- **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)**: Inline non-verbal tags ([laughter], [sigh] and a fixed list) and voice design by attributes are documented, but there is no emotion or pace direction and no multi-speaker request; each speaker is a separate call.
- **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))**: A per-utterance instruct steers tone, emotion and pace in plain language with InstructTTSEval scores on the card, but there are no inline tags, no speaker labels in one request and no seed parameter.
- **Resemble AI Text-to-Speech API (Resemble AI, Inc.)**: SSML is accepted but the supported tags are not listed, there are no speaker labels or dialogue mode, and the seed appears only in the response, not as a request field.
- **Soniox Text-to-Speech (Soniox Inc.)**: Bracketed audio tags for emotion, volume, pace and pauses are documented and must be written in English; unrecognised tags may be read aloud, one voice per stream means no labelled speakers in one request, and no seed exists.
- **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)**: Inline cues such as [cười] and [thở dài] are documented as experimental, style is ignored on v3 Turbo, and multi-speaker conversation exists in the web UI rather than as one API request.

- not this job: 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): No emotion tags, instruction field or speaker tags are documented; one voice per request.
- not this job: Deepgram Text-to-Speech (Aura-2, Flux TTS) (Deepgram, Inc.): Only a speed parameter, a beta expressivity dial from -2 to 2 and batch-only pause markers; no emotion tags, no speaker labels, and SSML or other vendors' tags are stripped.
- not this job: Rime TTS API + MCP (Rime Labs, Inc.): No emotion or style tags, no speaker labels and no seed; the only inline controls are speed and Mist v3 pauses.
- not this job: Smallest.ai Lightning TTS (Smallest, Inc.): No delivery tags, SSML or speaker labels on the TTS request; speed is the only delivery parameter.

## What stops agents finishing

- tags from one model family read aloud as text on another (Fish s1 parentheses on S2)
- pause and ellipsis tags triggering runaway generations (acc-31)
- second speaker rendered in the first speaker's voice (acc-32)
- direction only through SSML the voice ignores (Azure HD voices and personal voices skip some tags)

## Around it

Before: inference.chat-completion. After: video.compose-with-audio, text-to-speech.long-form-narration.

## Where this comes from

Weight 9 of 100 in Text-to-speech: 7 of 46 accounts tagged (acc-01, 08, 22, 31, 32, 38, 45); growing in 2026 as Drama 3, Inworld Reasoning, Gemini-TTS and ElevenLabs v4 ship direction, but most accounts are experiments

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