Prepare authorized references
Choose a clean speaker recording. Add a separate emotion reference only when the delivery needs emotional conditioning.
IndexTTS 2 · Expressive TTS · Emotion Control
IndexTTS 2 – Expressive AI Speech with Independent Emotion Control
IndexTTS 2 focuses on a part of text-to-speech technology that basic AI voice systems often struggle with: expression — separating speaker identity from emotional delivery so creators can shape how a voice performs without changing who it sounds like.
Online demo · Registered
0 / 1000
Spoken text · English · Sad
“I feel like I'm lost in the darkness and can't find a way out anymore.”
INDEXTTS 2 / 01 — Why IndexTTS 2
Why Expressive Speech Matters
Key idea
Traditional TTS systems often prioritize intelligibility.
If the words are pronounced correctly, the output is considered successful.
Creative media demands more.
A character delivering a dramatic line should not sound like a neutral navigation system.
An advertisement may require enthusiasm.
A reflective narration may need restraint.
A frightened character needs different delivery from the same character speaking calmly.
Expressive text-to-speech attempts to model these differences.
This is the problem IndexTTS 2 was designed to explore more deeply.
Duration-controlled autoregressive TTS Timbre-emotion disentanglement Zero-shot emotion + identity Expressive zero-shot TTS
Official English video example
Expressive speech generated with IndexTTS 2
An IndexTTS 2 demo demonstrating expressive, emotion-controlled speech.
INDEXTTS 2 / 02 — Emotion
Emotion is more than a voice style
IndexTTS 2 capability
Speaker Identity and Emotion Are Different
Key idea
One of the most important ideas behind the model is timbre-emotion separation.
Timbre describes the characteristics that make a speaker recognizable.
Emotion describes the way that speaker performs a particular line.
Those two properties often appear together in reference audio, but they are not the same thing.
IndexTTS 2 separates them conceptually.
A speaker reference can establish identity. A different signal can influence emotional delivery.
That makes the system much more flexible for creative applications.
IndexTTS 2 capability
Emotional Reference Audio
Key idea
One way to guide expressive generation is with emotional reference speech.
Imagine that you have a neutral recording representing the target speaker. You want the generated sentence to sound frightened.
Rather than requiring the same person to provide a frightened reference recording, the emotional direction can come from a separate example.
The generation system can then use one reference for identity and another for performance.
This is particularly useful when existing reference material does not contain every emotion required by a project.
INDEXTTS 2 / 03 — Emotion control
Emotion Vectors
Key idea
Another way to work with expressive speech is through structured emotion controls.
Instead of relying only on one reference clip, supported workflows can represent emotional direction using several dimensions.
This allows more deliberate control over the performance.
For example, a line may combine calmness with melancholy or surprise with fear.
This type of control can be useful for character dialogue and storytelling where emotions are not always easily reduced to one category.
Emotional reference audioEmotion intensity controlMulti-dimension emotion vectorText-guided emotional guidance
INDEXTTS 2 / 04 — Conditioning
Reference audio and emotion intensity
IndexTTS 2 capability
Emotion Intensity
Key idea
Emotion is rarely all or nothing.
A person may sound mildly disappointed rather than devastated. They may be enthusiastic without shouting. They may be slightly nervous rather than terrified.
The IndexTTS 2 workflow supports controlling how strongly emotional conditioning influences generated speech.
This is useful because subtle emotion is often more convincing than exaggerated performance.
Creators can compare different levels of intensity until the result matches the scene.
IndexTTS 2 capability
Text-Guided Emotional Speech
Key idea
Creators may also know the desired emotion without having suitable emotional reference audio.
Text-based guidance provides another way to communicate that intent.
A user might want “calm but concerned,” “excited and surprised,” “restrained anger,” or “soft and reflective.”
Natural-language guidance can help bridge the gap between creative direction and speech generation.
This makes expressive TTS easier to use in situations where a creator understands the performance they want but does not have a matching voice sample.
INDEXTTS 2 / 05 — Text-guided emotion
Speech Duration Control
Key idea
The other major research direction associated with IndexTTS 2 is speech duration.
Autoregressive text-to-speech models generate audio step by step, which makes the final length difficult to predict precisely.
For normal narration, this may not matter. For audiovisual production, it can become important.
A line may need to finish before a cut. Dubbed dialogue may need to fit inside an existing scene. An animation may provide a fixed speaking window.
IndexTTS 2 research introduces a method for bringing greater duration control into autoregressive speech synthesis.
INDEXTTS 2 / 06 — Duration control
Duration-controlled speech synthesis
IndexTTS 2 capability
Free Generation vs Controlled Duration
Key idea
Speech generation does not always need exact timing.
Sometimes the goal is natural prosody. At other times, timing matters more.
The research around IndexTTS 2 distinguishes between these situations.
One mode can target a predetermined speech length. Another can allow the model to generate more freely so prompt prosody influences the result.
This creates a useful conceptual tradeoff: natural unconstrained delivery vs tighter timing control.
That distinction is particularly relevant to dubbing and media synchronization.
Official English audio example
Duration-controlled English speech
“The equipment needed to do this includes rock saws and polishers.”
IndexTTS 2 capability
Why Duration Matters for Creative Media
Key idea
Timing affects how speech fits a scene.
Imagine translating a line from one language into another. The translated sentence may naturally take longer to speak. If the scene length stays unchanged, the voiceover may overlap the next shot.
A duration-aware workflow can help creators solve this type of problem.
The same issue appears in:
Animation
Dialogue replacement
Short-form videos
Product demos
Narrated presentations
Interactive characters
This is why duration research is an important part of the model even though ordinary TTS users may not need exact timing.
INDEXTTS 2 / 07 — Important note
Voice Cloning in an Expressive System
Key idea
Zero-shot voice cloning remains important, but in this generation it serves a larger purpose.
The key question is not simply: “Can the model imitate a reference voice?”
It becomes: “Can the model preserve that identity while changing how the speaker performs?”
That difference is central to understanding IndexTTS 2.
Voice cloning provides the identity foundation. Emotion controls provide the performance layer.
Together, they create a more flexible speech-generation system.
INDEXTTS 2 / 08 — Speaker identity
IndexTTS 2 for Characters
Key idea
Character performance is one of the clearest use cases for expressive speech.
A believable character changes emotion throughout a story — calm, then frustrated, then frightened, then relieved.
The voice identity should remain stable even though the performance changes significantly.
The separation of timbre and emotion makes IndexTTS 2 particularly interesting for game characters, animation, audio drama, virtual characters, interactive stories, AI companions, and narrative prototypes.
The value comes from maintaining continuity while varying performance.
INDEXTTS 2 / 09 — Use cases
Character performance and dramatic narration
IndexTTS 2 capability
Dramatic Narration
Key idea
Narration can also benefit from expressive control.
A single narrator may move through different emotional states during a story.
An introduction might sound confident. A tragic moment may become restrained and reflective. A climax may require more intensity.
Rather than replacing the narrator with a completely different synthetic voice, expressive generation can preserve identity while changing delivery.
This can produce more coherent long-form storytelling.
IndexTTS 2 capability
Character Performance
Key idea
Character performance is one of the clearest use cases for expressive speech.
A believable character changes emotion throughout a story — calm, then frustrated, then frightened, then relieved — while the voice identity stays stable.
The separation of timbre and emotion makes IndexTTS 2 particularly interesting for:
Game characters
Animation
Audio drama
Virtual characters
Interactive stories
AI companions
Narrative prototypes
The value comes from maintaining continuity while varying performance.
INDEXTTS 2 / 10 — Compare versions
IndexTTS 2 vs other generations
IndexTTS 2 capability
IndexTTS 2 vs Original IndexTTS
Key idea
The original IndexTTS focuses on controllable zero-shot TTS, speaker conditioning, pronunciation, and stable voice cloning.
IndexTTS 2 changes the center of attention.
Its defining ideas are:
Emotional speech
Timbre-emotion separation
Emotion intensity
Multi-source emotional guidance
Duration-control research
That makes the two versions complementary rather than repetitive.
IndexTTS 2 capability
IndexTTS 2 vs IndexTTS 2.5
Key idea
The difference between these two pages should also remain clear.
IndexTTS 2 is primarily about expressive speech and timing research.
IndexTTS 2.5 places more emphasis on multilingual generation, pronunciation, speed control, faster inference, and production-oriented efficiency.
Choose IndexTTS 2 when your project is centered on emotional delivery and character performance.
Explore 2.5 when multilingual production and newer workflow controls are more important.
INDEXTTS 2 / 11 — How to use
How to Use IndexTTS 2 for Expressive Speech
Generate a short baseline
Start with a short English script and neutral settings. Confirm speaker similarity before adjusting emotion intensity or text guidance.
Refine one control at a time
Change emotion, pacing, or pronunciation separately, compare each result, and keep the strongest settings for the final script.
INDEXTTS 2 / 12 — FAQ
IndexTTS 2 FAQ
IndexTTS 2 is an expressive zero-shot text-to-speech model focused on speaker identity, emotional control, and duration-aware speech-generation research.
INDEXTTS 2 / 12 — Explore
Explore Expressive Speech with IndexTTS 2
IndexTTS 2 moves AI speech beyond simple voice replication.
Its focus on emotional expression, identity preservation, and speech timing makes it particularly relevant for storytelling, characters, dialogue, and creative media.
Use it when the goal is not only to make text speak, but to make the generated voice perform.
References: GitHub · index-tts/index-tts · Official IndexTTS 2 page · IndexTTS 2 Technical Report (arXiv)