Niches · Text-to-speech

Speak an LLM reply to a live caller as it streams, and stop when interrupted

Two or three companies' docs show a full fit.

2companies, full fit
19companies, fit with a limit
49companies carry the tag
22share of the category's work
15first-hand accounts

What counts as done

A connection that lasts a 10-minute call, a cancel message that stops output, and only text sent before the cancel billed.

Measured, for our own tests: first audio within 300 ms of the first text at the median, audio keeps pace with the text so there is no gap mid-turn, the cancel stops output within one chunk, and no stream silently stops over the call. No docs we read state these figures, so they don't count against a listing here.

How agents ask for it: say this to the caller as soon as you can, turn my agent's reply into speech on the call, stream the LLM output to a voice and cut it off if the user talks

Does the job 2

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

ListingWhy
Deepgram Text-to-Speech (Aura-2, Flux TTS) A
Deepgram, Inc.
/v2/speak streams Speak tokens into a turn, Flush ends it with SpeechMetadata and a billable_character_count, Interrupt with playback_offset cancels and returns text_spoken, mulaw at 8 kHz is an output, and sessions run an hour; concurrency is 15 in the US and 5 elsewhere. source
Inworld Realtime TTS-2 C
Theai, Inc. dba Inworld AI
WebSocket contexts take LLM fragments with autoModeStrategy SENTENCE_BOUNDARY, flushContext, closeContext and cancelContext; contextClosed is the last message and the billing page says text not yet processed at a cancel is not counted; mu-law and A-law output. acc-13 runs it live and calls it fast. source

Does the job, with a limit 19

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
Amazon Polly BB
Amazon Web Services, Inc.
StartSpeechSynthesisStream (generative engine only) takes TextEvents as the LLM produces them, with a Force flush and a StreamClosedEvent, but there is no cancel message, the stream closes after 10 minutes or 5 idle seconds, only 8 run at once, and generative voices return no speech marks. source
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
The SGLang Omni adapter serves an OpenAI-compatible /v1/audio/speech with stream true, but it is pinned to one SGLang Omni commit, takes whole text per request with no cancel, and the card says streaming and throughput were not benchmarked for this release. source
Azure AI Speech text-to-speech BB
Microsoft Corporation
Input text streaming over the websocket v2 endpoint takes LLM chunks as they arrive, but only through the C#, C++ and Python SDKs, without SSML, with no documented cancel message or billing note, 10 minutes of audio per request, and 429s when a voice's regional capacity is busy. source
Cartesia Sonic TTS API + MCP B
Cartesia AI, Inc.
WebSocket contexts take fragments with continue, max_buffer_delay_ms and a done flag, but cancel only halts requests that have not started generating, Free allows 2 concurrent contexts, acc-29 reports request timeouts cutting calls and acc-30 clipped Japanese endings from collapsed timestamps. source
ElevenLabs Text to Speech API + MCP BB
Eleven Labs Inc.
The stream-input WebSocket takes text chunks with chunk_length_schedule, try_trigger_generation, flush and a close message, and the multi-context socket closes one context at a time, but no cancel-billing note exists and acc-04 found real-time latency well above the advertised figure; acc-05, acc-36 and acc-37 run agents on it. 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
SGLang-Omni serves a streaming /v1/audio/speech with 89 ms median first audio at one stream on an H200 by Fish Audio's own table, but the request takes whole text, cancelling means closing the connection, 48 GB of NVIDIA memory is needed, and acc-31 saw 6 to 20 percent runaway takes. 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
The api_server streams audio with --streaming and the card claims about 100 ms to first audio on an H200, but the request takes whole text, there is no cancel message, 24 GB of GPU memory is asked for, and the Research License forbids commercial use. source
Fish Audio TTS API B
Hanabi AI Inc.
The WebSocket takes text fragments with flush and stop events and ends with a finish event, but there is no cancel message (stop drains the buffer), billing after a cut is unstated, the 300 ms first audio is Fish Audio's claim, and concurrency is 5 below USD 100 prepaid. source
Google Cloud Text-to-Speech B
Google LLC
StreamingSynthesize takes a stream of text inputs after one config and returns audio as it goes, with MULAW output, but only for Chirp 3 HD voices, with no flush, cancel or done message documented and no note on billing after an early close; billing must be enabled first. source
HeyGen Voice BB
HeyGen Technology, Inc.
wss stream-input takes LLM text in pieces with flush, close_context, isFinal and synthesizedChars, and ulaw at 8 kHz is an output option, but 30 requests a minute per member counts every handshake and context, Turbo Free returns 429 on shared capacity, and no latency figure is published. 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)
vLLM-Omni serves /v1/audio/speech with PCM streaming out, but each request needs the whole text, no cancel or done semantics are documented, there are no timestamps, and the recipe targets one H100 or H200. source
MiniMax Speech (T2A) C
Nanonoble Pte. Ltd.
The bidirectional WebSocket takes task_continue at any granularity, buffers to sentences, and has task_flush, task_cancel (keep sending afterwards) and is_final with usage_characters, but T2A is capped at 60 requests a minute, no latency figure is given, and the docs do not say what a cancelled sentence costs. source
Murf TTS API + MCP BB
Murf Inc.
The Falcon 2.1 WebSocket takes text in chunks per context_id with end, a clear frame on interruption, a final frame and word timestamps, and streams 8 kHz ULAW; but concurrency is 2 outside US-East (5 there), the 100 ms figure is Murf's own, and billing after clear is unstated. 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)
The vLLM-Omni recipe serves /v1/audio/speech with stream true and PCM, but the qwen-tts package has no streaming or server, each request needs the whole text, no cancel exists, and acc-07 builds its own VAD and streaming pipeline around it. source
Resemble AI Text-to-Speech API D
Resemble AI, Inc.
The WebSocket streams audio per request with request_id and a terminal audio_end, but it is Business plan only, each message is a whole text of up to 3,000 characters rather than LLM fragments, and no cancel message exists; the agent must close the socket to stop. source
Rime TTS API + MCP C
Rime Labs, Inc.
Coda WebSocket v1 streams LLM tokens into a context with SPLIT_STRATEGY_ACCUMULATE_SENTENCE, end, stop and cancel, answering cancelled or done, at 8 kHz PCM or mu-law; but Rime publishes no concurrency figure and no rule on what a cancelled context costs, and the 96 ms P50 figure is Rime's own. source
Smallest.ai Lightning TTS B
Smallest, Inc.
WebSocket takes LLM fragments under one context_id, cancel_request discards buffered text, 8 kHz ulaw output and word timestamps exist, but the concurrency page says 1 request per account, no frame marks a context done, and billing after a cancel is not stated. source
Soniox Text-to-Speech BB
Soniox Inc.
The WebSocket takes text chunks with text_end, a cancel message stops audio, audio_end and terminated close the stream and 8 kHz output exists, but 3 concurrent streams by default, a stream dies after a few seconds without text, and billing after a cancel is not stated. 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
The OpenAI-compatible server streams pcm or wav chunked or over SSE with an 8 kHz resample option, and the author measures 105 to 115 ms first audio on an RTX 3060, but the request takes whole text, cancelling means closing the connection, and it speaks only Vietnamese and English. source

Doesn't do this job 2

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

ListingWhy
Ming-3.0-tts-0.5B C
inclusionAI (the LICENSE's copyright holder; source headers read Copyright (c) Ant Group)
The publisher's infer.py returns one whole waveform; no streaming, no text input over a socket, no cancel and no server are documented for the 0.5B checkpoint. 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
generate() returns whole waveforms; no streaming, no server endpoint and no cancel in the package, only a Gradio demo and CLI tools. source

What stops agents finishing

  • no WebSocket that takes text as it is produced (REST only, or WebSocket only on a Business plan as at Resemble)
  • concurrency of 2 to 5 streams on the plan an agent can sign up to (Cartesia Free 2, Soniox 3, Murf 2 outside US-East)
  • request timeouts with no fallback provider, which leave the caller in silence
  • end of speech only estimated from text, so the call ends before the last word plays

Around it

Before: speech-to-text.voice-agent-live, inference.chat-completion. After: voice-agents.play-to-call, telephony.hang-up.

Where this comes from

Weight 22 of 100 in Text-to-speech: 15 of 46 accounts tagged (acc-03, 04, 05, 06, 07, 11, 13, 18, 28, 29, 30, 36, 37, 41, 44); the job every API vendor leads with and the one with the most failure reports, so weighted at the top

The same as data: /api/v1/use-cases/text-to-speech.voice-agent-turn.json, or the MCP tool get_use_case with id text-to-speech.voice-agent-turn. 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.