Two local Vietnamese text-to-speech models against edge-tts

2026-11-14

For my project I wanted Vietnamese speech that was faster or better than edge-tts, the free tool that calls Microsoft's online voices. At the end of July 2026 I tried two local models against it. Both lost on the numbers that matter for Vietnamese.

How I tested

Same machine and run as its own edge-tts control in each test. I measured:

RTF (real-time factor): seconds to generate divided by seconds of audio. Lower is faster.
WER (word error rate): the share of words in the full transcript that came out wrong.
Tone error, in the VoxCPM2 test only: errors on Vietnamese tone marks.

supertonic, 31 July 2026

edge-tts (control): Audio 20.0 s; Generate 2.8 s; RTF 0.14; WER 9.9%
supertonic, 8 steps: Audio 40.4 s; Generate 6.5 s; RTF 0.16; WER 22.0%

Twice the word error rate, and not faster. Changing the step count didn't trade quality for speed, the output just fell apart. At 4 steps WER was 35.2%, at 2 steps 100%. 15 of the 40.4 s was silence; I did not tune its speed or silence settings.

VoxCPM2, 30 July 2026

edge-tts, voice vi-VN-HoaiMyNeural: Audio 20.0 s; Generate 1.1 s; RTF 0.05; WER 7.7%; Tone error +0.31%
VoxCPM2 (PyTorch on Apple's MPS): Audio 23.4 s; Generate 15.4 s; RTF 0.66; WER 9.9%; Tone error +1.69%

VoxCPM2 does sound good to the ear. It was also about 13 times slower to generate (RTF 0.66 against 0.05), worse on word errors, and five times worse on tone marks. In Vietnamese a wrong tone mark changes the word, so that is the error I care about most.

Don't compare across the two tables

The edge-tts control scored 9.9% WER in one test and 7.7% in the other, for a reason I did not record (the run, voice or text differed). Put supertonic's 22.0% next to VoxCPM2's 9.9% and you get a confident, wrong ranking. Compare each model only with the control from its own run.

What I did not test

One script per test and one run each. I did not try other voices or longer scripts. Development had already ended on 23 July 2026, a week before my test; the repo is being archived (see its tool page). I kept edge-tts.