Niches · Text-to-speech

Read names, numbers, acronyms and codes precisely

Four or more companies' docs show a full fit.

7companies, full fit
13companies, fit with a limit
38companies carry the tag
10share of the category's work
7first-hand accounts

What counts as done

A year, a unit, a six-character code, acronyms and surnames are handled in the request language by the model's documented normalisation or by a documented override (dictionary, phoneme tag, alias) applied in one further call.

Measured, for our own tests: on a fixed test list (a year, a unit such as 135 ms, a six-character code, three acronyms, two surnames) every item is read correctly in the request language, or the override fixes a miss in one further call. No docs we read state these figures, so they don't count against a listing here.

How agents ask for it: read back the account number, say this product name right, speak this confirmation code

Does the job 7

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

ListingWhy
Amazon Polly BB
Amazon Web Services, Inc.
say-as, sub and phoneme tags plus PLS lexicons (alias or phonetic, up to five per request, also on the stream via LexiconNames) fix a miss in one further call; phoneme is partial on generative voices. source
Fish Audio TTS API B
Hanabi AI Inc.
Pronunciation dictionaries with phoneme and alias rules are created and versioned over the API or sent inline, phoneme tags cover English, Chinese and Japanese, and normalize plus a language tag spells out numbers and units in 15 language families, on the WebSocket too. source
Google Cloud Text-to-Speech B
Google LLC
SSML say-as (cardinal, characters, verbatim, telephone), sub and phoneme work on sync requests, and Chirp 3 HD takes a custom_pronunciations field in IPA or X-SAMPA; SSML is not applied on streaming requests. source
Inworld Realtime TTS-2 C
Theai, Inc. dba Inworld AI
Inline /IPA/ replaces a word in the text, a workspace pronunciation dictionary applies with enable_custom_pronunciation on TTS-2 requests with an explicit language, and the docs index lists verbatim tags for codes and IDs. source
MiniMax Speech (T2A) C
Nanonoble Pte. Ltd.
pronunciation_dict tone rules replace a word with plain text, pinyin, IPA or kana on sync, async and WebSocket routes, and inline (IPA) or (pinyin) in parentheses overrides one occurrence, so a miss is fixed in one further call. source
Murf TTS API + MCP BB
Murf Inc.
Inline IPA in quotes and slashes in the text, or pronunciationDictionary with IPA or SAY_AS entries, work on Stream Speech and WebSocket voice_config on Falcon 2.1, so a wrong surname is fixed in one further call. source
Rime TTS API + MCP C
Rime Labs, Inc.
Normalisation expands numbers, dates, currency and units, spell() forces letter-by-letter codes on both models, customLexicon sends X-SAMPA pronunciations with Coda requests over HTTP, WebSocket and gRPC in six languages, and /textnorm shows what will be read. source

Does the job, with a limit 13

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.

ListingWhy
Azure AI Speech text-to-speech BB
Microsoft Corporation
phoneme (IPA, SAPI, UPS), say-as, sub and custom lexicons are documented, but HD voices support only a subset of SSML, and acc-27 shows an en-US multilingual voice reading numbers in Chinese with Microsoft offering only SSML or rewording as a fix. source
Cartesia Sonic TTS API + MCP B
Cartesia AI, Inc.
Pronunciation dictionaries (IPA or sounds-like) apply on every TTS endpoint through pronunciation_dict_id, a spell tag forces character read-out and a normalization field sets number conventions, but entries default to case-insensitive on Sonic 3.6, the September change the listing records. source
Deepgram Text-to-Speech (Aura-2, Flux TTS) A
Deepgram, Inc.
Inline IPA pronunciation is Early Access on Flux, English only and cannot be combined with speed or pause; Aura-2 has it in English and Spanish; numbers and dates are normalised in ten languages, the rest must be written out; SSML is stripped. source
ElevenLabs Text to Speech API + MCP BB
Eleven Labs Inc.
Pronunciation dictionaries with alias and IPA or CMU phoneme rules apply through pronunciation_dictionary_locators (up to three), v4 reads inline /IPA/, and apply_text_normalization has on, off and auto; but phoneme rules work only on eleven_flash_v2, Flash v2.5 misreads large numbers, and its normalisation toggle is Enterprise only. source
Fish Audio Drama 3 C
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
Fish Audio's Drama 3 page documents phoneme tags (Arpabet, pinyin, romaji) as a model feature, but the fish-speech inference guide shows no dictionary, alias or normalisation option for the local server, so a miss is fixed only by editing the text. source
Fish Audio S2 Pro D
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
Fish Audio's docs describe phoneme tags as a feature of the S2 family, but the fish-speech inference guide for the local server lists no dictionary, alias or normalisation control, so a misread must be fixed by rewording. source
HeyGen Voice BB
HeyGen Technology, Inc.
pronunciation_dictionary takes up to 100 IPA terms on both speech endpoints, but only on heygen-voice-1 and heygen-voice-1-turbo (not Turbo Free), matching ignores case, and there is no number or say-as control. source
Ming-3.0-tts-0.5B C
inclusionAI (the LICENSE's copyright holder; source headers read Copyright (c) Ant Group)
FST text normalisation covers numbers, dates, units and symbols but needs pynini, which has no universal wheel, and the README's first place on Artificial Analysis pronunciation robustness is self-reported; no dictionary or phoneme override exists. source
Ming-3.0-tts-tiny C
inclusionAI, Ant Group (named in the site footer of inclusion-ai.org; the LICENSE copyright line reads inclusionAI)
FST normalisation for numbers, dates and units needs pynini, the Artificial Analysis pronunciation lead in the README is self-reported, and no dictionary or phoneme override exists. source
OmniVoice C
k2-fsa, the Next-gen Kaldi open-source project. The paper's authors are at Xiaomi Corp.; no company is named as the licensor
English pronunciation is overridden inline with CMU Arpabet in brackets and Chinese with tone-number pinyin, but digits are read digit by digit unless normalize_text (an optional extra) is on, and other languages fall back to num2words for integers only. source
Resemble AI Text-to-Speech API D
Resemble AI, Inc.
data accepts text or SSML on all three routes, with partial SSML support on HTTP streaming, but no page lists which tags (phoneme, say-as, sub) Resemble Ultra honours, so an override cannot be confirmed. source
Smallest.ai Lightning TTS B
Smallest, Inc.
pronunciation_dicts fix names and acronyms on sync, SSE and WebSocket routes, but entries are respellings, not IPA, there is no SSML, and nothing documents how years, units or codes are read. source
Soniox Text-to-Speech BB
Soniox Inc.
Soniox claims accurate alphanumerics, phone numbers and IDs, but no dictionary, phoneme tag or alias is on either API, so a miss cannot be fixed without rewording. source

Doesn't do this job 3

Listings in the category that do something else, with the reason.

ListingWhy
Audio8-TTS-Realistic 0.6B D
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 phoneme tag, dictionary, alias or normalisation option in the card or README. source
Qwen3-TTS 12Hz 1.7B CustomVoice C
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)
No pronunciation dictionary, phoneme tag, SSML or normalisation setting is documented; the only control is a natural-language instruct for style. source
VieNeu-TTS v3 Turbo C
Phạm Nguyễn Ngọc Bảo, an individual (named on the card as author; Hugging Face account pnnbao-ump, GitHub account pnnbao97). No company found
No phoneme tag, dictionary or alias is documented; the g2p phonemizer runs internally with no override. source

What stops agents finishing

  • no pronunciation control at all (Deepgram strips SSML; ElevenLabs Flash v2.5 number normalisation is Enterprise only)
  • a dictionary that silently changes behaviour (Cartesia dictionaries became case-insensitive in September 2026)
  • LLM-written text the agent cannot pre-edit, so a fix that needs rewording (acc-27) does not apply

Around it

Before: inference.chat-completion. After: text-to-speech.voice-agent-turn.

Where this comes from

Weight 10 of 100 in Text-to-speech: 7 of 46 accounts tagged (acc-08, 09, 10, 24, 26, 27, 42); every one is a failure report, and the buyer check names it, so weighted above its count

The same as data: /api/v1/use-cases/text-to-speech.names-numbers-codes.json, or the MCP tool get_use_case with id text-to-speech.names-numbers-codes. Other jobs in this category: /api/v1/niches?category=text-to-speech.

Draft, read 2026-11-12 from public docs and first-hand accounts, not yet reviewed by a person. A mark says what the docs show, not what a run measured.

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.