Inworld 发布 Realtime TTS-2 更新：引入推理功能并优化性能

Today we're shipping two updates to Realtime TTS-2. Reasoning makes it the first speech model that decides how to say a line before saying it: TTS-2 reads the text and any context you send, writes a short delivery direction, then performs it. An updated model generates speech faster, matches reference voices more closely and holds steady across long sessions and long text.

Both run behind the same public model ID, inworld-tts-2. The update reaches existing integrations automatically as it rolls out by serving region; Reasoning is opt-in, one field on the request.

Update 1 of 2

Reasoning

Realtime TTS‑2 reads a line and its context, decides how it should sound, then performs it.

01How it works

Speech models read the text they're given, and nothing tells them how it should sound. In a consumer app, your LLM generates every reply at runtime, so there's no one to add that direction by hand. A line like "I can't believe you did that." can mean delight, hurt or disbelief, and without context the model gives each of them the same neutral read.

Reasoning adds one step before synthesis, inside the same TTS-2 request:

There's no extra service to call or host. The reasoning step is a language model whose only job is to write the delivery direction for TTS-2. The direction comes back in the API response, and any instructions or tags you set yourself take precedence.

02Why it beats emotion tags

Emotion tags are a fixed list. Something has to pick a label like [happy] for each line, usually your own LLM following prompt rules, and the voice plays that label. Reasoning writes a free-form direction instead, like warm, delighted, speaking through a smile, so a single read can set emotion, tone and pacing together, based on the context you pass.

Saves time

No tag vocabulary in your prompt, no rules to maintain, no scripts to tag by hand. Enable it and pass the context you already have.

Sounds more natural

Each line gets a direction based on its context, so delivery varies the way a person's does instead of switching between a few preset emotions.

03Try it today

Enable Reasoning in the request and pass context in its own field. Context shapes the delivery but is never spoken. The response includes the generated direction, so you can log it alongside the input and adjust your context when a read is off.

Request

curl -X POST https://api.inworld.ai/tts/v1/voice \
  -H "Authorization: Basic $INWORLD_API_KEY" \
  -H "Content-Type: application/json" \
  -d @- <<'JSON'
{
  "text": "I can't believe you did that.",
  "voiceId": "Reed",
  "modelId": "inworld-tts-2",
  "synthesisContext": {
    "reasoning": "REASONING_ON",
    "reasoningContext": "A friend has arranged an unexpected reunion."
  }
}
JSON

Response

{
  "audioContent": "UklGRiQAAABXQVZFZm10...",
  "usage": { "processedCharactersCount": 29, "modelId": "inworld-tts-2" },
  "reasoningInstruction": "Breathless and stunned, like a friend seeing a long-lost face."
}

Reasoning runs on TTS-2 synthesis, streaming and async requests. That suits audio you generate per message or ahead of time, like voice messages from a companion, generated stories and lessons, audiobooks, or pre-generated character lines. WebSocket streaming, duplex (including the Realtime API) and batch synthesis are coming soon.

For parameters and limits, see the Reasoning documentation.

Update 2 of 2

An updated model

Faster generation, closer voice matches and steadier long sessions, behind the same model ID at the same price.

04Faster generation across utterance lengths

The updated model addresses the failures that become apparent across real workloads. A voice application can move from a short acknowledgment to a detailed explanation, then read a date or confirmation code aloud. A companion might continue that exchange for an entire session. Across those, a voice can gradually change, a word can be skipped in a long passage, or a consonant lost at the end of a sentence.

In our comparisons with the previous version of TTS-2, short utterances are about 13% faster to generate, and mid-length utterances are about 9% faster. Generation speed is also more consistent across input lengths. That matters for applications whose responses vary in length: a voice agent may acknowledge a request in a few words before giving a longer explanation, and a tutor may move between a quick correction and a full worked example.

These figures describe TTS generation performance. The response delay a listener experiences also depends on turn detection, language model generation, network transit, and playback buffering. Measure those components separately when evaluating your application's latency.

05More consistent voices and more reliable speech

The updated model improves voice similarity in Stable mode, particularly for voices that have been difficult to clone accurately, and reduces voice drift during long duplex sessions. Custom voices stay closer to the reference speaker, while companions and voice agents maintain a more consistent identity as the conversation develops.

It also fixes cases where TTS-2 skipped words in long utterances or dropped plosives on some word endings. Preserving these consonants, such as /p/, /t/, and /k/, makes affected word endings clearer. For audiobooks and articles read aloud, complete readings preserve the source text; for language learners and callers, clear pronunciation makes spoken details easier to understand.

Creative mode now produces more natural delivery, with light built-in steering. For interactive characters and companions, this improves how a line is performed as well as how the voice sounds. Use Stable mode when voice consistency is the priority, and test Creative mode where a more expressive performance fits the experience. The audio generation documentation explains the available delivery modes.

Evaluate these improvements across full passages and extended sessions, with the lines your application actually generates. Compare cloned voices with their references, listen to samples from the beginning and end of a conversation for drift, and check generated speech against the source for omitted words or unclear endings. Assess these separately from naturalness and expressiveness.

06Independent benchmarks for speed and accuracy

Cekura's independent benchmark provides context for the TTS-2 family. Its results, updated October 28, 2026, distinguish TTS-2 from Realtime TTS-2 Flash:

• Realtime TTS-2 Flash: 46 ms median, 73 ms p90 to the first audible sound, with the complete text supplied upfront. Cekura ranks that the fastest of the 27 models it tested in the non-streaming category. Its word-by-word input test is a separate measure.
• Realtime TTS-2: 93.4% of difficult items read correctly, across numbers and shorthand, codes and addresses, words that depend on context, and medical terms.

Cekura times the first result from the moment the full sentence is sent to the first audible sound. The accuracy result is estimated from speech-to-text transcripts rather than human listening. Every model read the same 104 agent sentences, three times each, from one US East cloud region; see the methodology. These are results for the models Cekura tested, before this update and without Reasoning. The 13% and 9% speed improvements above come from our comparison of the updated and previous models.

07Keep your integration and hear the difference

If you already use inworld-tts-2, continue using your existing integration. The update rolls out behind that identifier, with no API migration and no price change. Reasoning is opt-in: nothing changes in your audio until you enable it on a request.

Try the updated model on the inputs that matter to your product: a difficult reference voice in Stable mode, an extended duplex conversation, or a long passage that previously lost words. If you have saved baseline audio, use the same text, voice, and settings to make the comparison useful. Then enable Reasoning on a few lines with the context your app already has, and compare the reads.

For a repeatable evaluation, the Inworld TTS Open Evaluation Toolkit provides per-sample measurements and reports that let you inspect the errors behind aggregate scores. Combine transcript checks with listening tests on your production voices and languages.