Idu Mishmi · Text-to-speech evaluation

Does the machine sound like the speaker?

Three Idu Mishmi VITS models, one set of 20 held-out sentences. Listen to each model beside the original recording, and judge for yourself.

01

Overview

We trained three versions of an Idu Mishmi voice and asked each to read the same 20 sentences that none of them saw during training. Every clip is shown unedited, next to the original speaker.

  • Original the real recording of the sentence.
  • Models A, B, C synthesized from the sentence text only.
  • Blind mode hides model names and shuffles order, so you listen without bias.
02

The three models

Clip length vs the original

Average difference between the generated clip's length and the original recording's, across all 20 sentences. Lower means pacing closer to the real speaker. This is a pacing indicator only, not a quality score.

03

Listen & compare

04

All 20 sentences

05

About this comparison

Test set

20 common held-out sentences, identical for every model. Audio is 22.05 kHz mono WAV. All clips are shown without post-processing.

Reading the results

This page is for listening. It does not yet include listener scores (MOS) or preference votes. Use blind mode and compare pronunciation, rhythm and naturalness against the original.