Inworld TTS-2 Adapts Its Voice to the Tone of a Conversation

Inworld has released Realtime TTS-2, a voice model that follows the audio context of a conversation, accepts performance directions, and speaks more than 200 languages.

The sentence remains the same, but its delivery changes according to what just happened. After a joke, the voice responds more lightly. When speaking to someone who sounds hesitant or upset, it slows down and lowers its energy. Realtime TTS-2 aims to make speech synthesis an attentive participant in a conversation rather than a simple text reader.

Inworld is making its new model generally available after several months in research preview. It can be accessed through the company’s text-to-speech API and its Realtime interface, which brings transcription, a language model, and audio output together within the same session.

The launch includes two versions. Realtime TTS-2 is the flagship model, focused on expressiveness and control over delivery. Realtime TTS-2 Flash prioritizes speed and cost for applications where a few dozen milliseconds can affect the flow of an exchange.

The difference from conventional speech synthesis is not limited to audio quality. A traditional service receives a sentence and produces its spoken equivalent without necessarily knowing what the user said earlier or how they said it. Realtime TTS-2 can take the audio from previous turns into account when the conversation runs through the Realtime API.

The system does not rely on the transcript alone. It receives signals drawn from the user’s voice, including pace, intonation, and inferred emotional state. These elements influence how the next response is delivered.

This continuity relies on a persistent connection. Each spoken turn becomes part of the session context, and the language model’s response is then passed to TTS-2. The voice layer decides how to perform that response based on previous exchanges and the developer’s instructions.

The different stages still need to be separated. Realtime TTS-2 does not necessarily choose the words in the response; that task belongs to the connected language model. Its main role is to render those words as speech. Total latency also depends on transcription, reasoning, tool calls, network conditions, and audio playback on the user’s device.

One of the main additions is natural-language voice direction. An instruction placed inside brackets can ask the voice to speak warmly, whisper, restrain its anger, or sound exhausted after a long day. The same voice and the same text can therefore produce several interpretations.

This method goes beyond a fixed list of emotions. Developers can specify an intention, a situation, or a shift within the sentence. Inworld recommends detailed descriptions rather than very short commands: “speak sadly, as if something bad just happened” provides more context than the word “sad.”

Directions can also be inserted during a line. The model supports nonverbal events such as laughter, sighs, breaths, throat clearing, coughing, and yawning. These are rendered as sounds rather than spoken as text.

TTS-2 can also introduce hesitations, self-corrections, and unfinished thoughts. These irregularities are intended to bring the delivery closer to spontaneous conversation. They will not suit every situation: customer support, audiobooks, and regulated announcements may require more stable speech.

Inworld therefore offers three behavior settings. Expressive allows greater variation and targets characters, virtual companions, and consumer conversations. Balanced serves as the general-purpose default. Stable reduces shifts in pitch and delivery for phone systems, longer content, and professional applications that require stricter consistency.

This ability to follow context could be useful for customer service, language learning, games, tutoring, and wellness applications. A virtual teacher could maintain a recognizable voice while adjusting its energy. A game character could deliver the same line differently depending on the tension of the exchange. A support agent could attempt to adopt a calmer tone when speaking to an irritated customer.

Adaptation should not be mistaken for certainty about a person’s emotional state. A quiet voice may indicate fatigue, sadness, a personal vocal trait, or simply a poorly configured microphone. Speaking pace and intonation also vary across languages, cultures, accents, and certain disabilities.

A mistaken interpretation could produce an inappropriate response. An artificially sympathetic voice may become irritating if the system has misunderstood the situation. In medical, administrative, or financial settings, vocal signals should therefore supplement explicit context rather than serve as the sole basis for conclusions about the user.

The multilingual feature is designed to preserve the same voice identity when a sentence changes language. The examples do not require a separate language flag: the model detects the transition and attempts to maintain the voice’s timbre, pitch, and character.

The official pages do not all describe the same coverage. The Realtime TTS-2 launch page says that one identity can be preserved across more than 100 languages. The main voice product page and pricing table advertise support for more than 200.

Inworld acknowledges that these languages do not all have the same level of maturity. The highest tier is presented as offering native-speaker quality, while part of the newly added long tail remains experimental. A language appearing on the list does not therefore guarantee pronunciation, prosody, and identity preservation equivalent to those available in English.

Switching languages within a single sentence is particularly demanding. Maintaining the same identity is not enough: accents, phonemes, and rhythms must change without making it sound as though a different speaker has taken over. Inworld has not yet published detailed results for every language or language combination.

The service also adds the ability to create voices from written descriptions. Advanced Voice Design accepts a few lines defining age, accent, energy, pitch, or performance style. It then produces a saved voice that can be reused in later API