Phonemic or graphemic input?

Devanagari does not write schwa deletion. नमक is spoken namak, not namaka, and nothing in the spelling says so. One arm below is handed the deletion by a rule-based front end. The other reads the characters and has to work it out. Everything else about the two is the same: architecture, seed, corpus, step budget. The question this page exists to answer is whether that front end still buys anything.

  1. Read the Devanagari sentence. The highlighted words are the ones where medial schwa deletion applies, so they are where the two arms can differ.
  2. Play Natural speech first. That is the human speaker, and it is the thing both arms are trying to be.
  3. Then play the two arms under the comparison and listen to the highlighted words only. Everything else is shared.
  4. The groups below that exist to tell you how much of any difference you hear is real: the same arm at other seeds, and the same arm trained on less data.

Sentences are from the held-out split, which no model trained on. Each sentence carries the natural recording, so the question can be how far each arm is from the speaker rather than only which arm you prefer.

How these sentences were chosen

Reading this page

Phonemic front end
A rule-based grapheme-to-phoneme converter runs first. It deletes the schwa that the spelling writes but nobody says, so the model is handed n ə m ə k as n ə m k and never has to learn the rule.
Graphemic, raw characters
No front end. The model reads न म क and has to infer the deletion from the audio it was trained on. This is the arm that tells you whether the front end is still doing work.
Rung
How much speech that run was trained on. The rungs are nested subsets of one speaker, so 1 h is inside 5 h is inside 9 h. The step budget is identical at every rung, which means a smaller rung sees its data more times, not less often.
Seed
The random seed of the training run. Two runs that differ only by seed differ only by training noise, so the spread between them is the scale any claimed difference has to beat.
Vocoder
FastSpeech 2 emits a mel spectrogram, which is not audio. A vocoder turns it into a waveform. VITS produces a waveform itself, and is marked end to end. Every clip says which it was, because that is the one thing you cannot hear and cannot recover.

What you are hearing