Resource workflow

Listen, decide, then test with your own material

Each resource has one job. Start with attributed examples, understand the model differences, then run a controlled comparison with your own authorized reference.

01 · Listen

Samples

Hear attributed official reference and showcase examples across supported languages.

02 · Decide

Model comparison

Compare hosted capabilities, upstream evidence, limitations, and model-selection guidance.

03 · Test

Voice Lab

Use your own authorized reference and script for a controlled IndexTTS2 vs 2.5 test.

Voice Lab

Compare IndexTTS2 and IndexTTS 2.5 with controlled inputs

Upload one reference, enter one comparison script, then generate each model from the same live workspace. Voice Lab keeps successful IndexTTS2 and IndexTTS 2.5 outputs so you can listen side by side without duplicating inputs.

Prepare one shared input

Upload one clean reference recording and enter the target text once.

Generate the baseline

Run IndexTTS2 first, or start with 2.5 if that better matches your workflow.

Switch models and compare

Keep the reference and text unchanged, generate the other model, then listen to both captured outputs below.

Shared comparison workspace

One reference. One script. Two model outputs.

The generator below is the same hosted product pipeline used elsewhere on IndexTTS Online. Switching models preserves the shared reference and text while the lab retains each successful output for comparison.

For a direct model A/B test, use English or Chinese because both hosted models support those target languages. Japanese, Spanish, and Arabic are IndexTTS 2.5 capability tests rather than symmetric A/B comparisons.

IndexTTS Online Voice Cloning

Online

Clone a voice, then generate natural speech

Add a short reference voice and the text you want it to speak.

0 / — characters

Model

Choose the IndexTTS model for this generation.

Reference voice

Add 3–30 seconds of clear, single-speaker audio.

Text to speak0 / —

Supports English and Chinese text.Punctuation helps control pauses and delivery.

Output

Waiting

Generated audio will appear here.

Add a reference voice and text to get started.

WAV output · Signed-in generations can be viewed in History.

Captured A/B outputs

Each successful generation is stored in this page session under its model. Regenerate a model to replace that side with the latest result.

Configuration A

Start with IndexTTS2

Use the same reference audio and target text as Configuration B for a fair comparison.

Generate IndexTTS2 in the workspace above to capture side A.

Configuration B

Start with IndexTTS 2.5

Keep the reference and text fixed, then change one supported control at a time when testing emotion, language, or pace.

Generate IndexTTS 2.5 in the workspace above to capture side B.

Fair-comparison checklist

Use the same reference clip and the same target text in A and B.
Use a reference with one speaker, little background noise, and no music.
Compare one factor at a time instead of changing model, emotion, and pace together.
Listen for pronunciation, speaker similarity, prosody, stability, and artifacts rather than relying on a single impression.
Repeat important comparisons because generative output can vary between runs.

What this lab does not claim

Voice Lab is a comparison workspace, not a published quality score. We do not invent MOS, speaker-similarity, or naturalness numbers. The benchmark page separates upstream published findings from IndexTTS Online product guidance.

Read the model comparison →Browse source samples →

Generations use the normal IndexTTS Online quota, billing, moderation, storage, and safety pipeline. Use only voice references you have permission to use. See our Terms and Acceptable Use Policy.

Practical guides

Learn the workflow behind the tool

Use focused guides to improve reference audio, choose models, prepare scripts, and compare results under controlled conditions.

9 min read

A Reproducible A/B Testing Method for AI Voice Cloning

Compare voice-cloning models, references, scripts, emotion, or pace with controlled inputs, useful listening criteria, repeat runs, and honest reporting without fabricated MOS.

Read guide

6 min read

How to Choose Better Reference Audio for AI Voice Cloning

A quality checklist for reference recordings, with practical fixes for noise, echo, multiple speakers, clipping, and unnatural delivery.

Read guide

6 min read

Voice Cloning Troubleshooting: Why Generated Speech Sounds Wrong and What to Fix First

Diagnose common voice-cloning problems by separating reference-audio issues, script issues, pronunciation, model choice, pacing, and output review.

Read guide

Browse all guides

Resource workflow

Listen, decide, then test with your own material

Each resource has one job. Start with attributed examples, understand the model differences, then run a controlled comparison with your own authorized reference.

01 · Listen

Samples

Hear attributed official reference and showcase examples across supported languages.

02 · Decide

Model comparison

Compare hosted capabilities, upstream evidence, limitations, and model-selection guidance.

03 · Test

Voice Lab

Use your own authorized reference and script for a controlled IndexTTS2 vs 2.5 test.

Compare IndexTTS2 and IndexTTS 2.5 with controlled inputs

Prepare one shared input

Upload one clean reference recording and enter the target text once.

Generate the baseline

Run IndexTTS2 first, or start with 2.5 if that better matches your workflow.

Switch models and compare

Keep the reference and text unchanged, generate the other model, then listen to both captured outputs below.

One reference. One script. Two model outputs.

IndexTTS Online Voice Cloning

Online

Clone a voice, then generate natural speech

Add a short reference voice and the text you want it to speak.

0 / — characters

Model

Choose the IndexTTS model for this generation.

Reference voice

Add 3–30 seconds of clear, single-speaker audio.

Text to speak0 / —

Supports English and Chinese text.Punctuation helps control pauses and delivery.

Output

Waiting

Generated audio will appear here.

Add a reference voice and text to get started.

WAV output · Signed-in generations can be viewed in History.

Fair-comparison checklist

Use the same reference clip and the same target text in A and B.

Use a reference with one speaker, little background noise, and no music.

Compare one factor at a time instead of changing model, emotion, and pace together.

Listen for pronunciation, speaker similarity, prosody, stability, and artifacts rather than relying on a single impression.

Repeat important comparisons because generative output can vary between runs.

What this lab does not claim

Generations use the normal IndexTTS Online quota, billing, moderation, storage, and safety pipeline. Use only voice references you have permission to use. See our Terms and Acceptable Use Policy.