Samples
Six clips from samples/, synthesized by this model on the Arm CPU alone
(ONNX Runtime, 9 threads) on a DGX Spark GB10. No GPU involved, no cherry
picking — these are the first six sentences of the held-out evaluation set in
order, the same set the WER/CER numbers in EVAL.md come from.
| file | text |
|---|---|
samples/sample_00.wav |
The weather changed suddenly this afternoon. |
samples/sample_01.wav |
Could you please pass me that book? |
samples/sample_02.wav |
Wow, I did not expect to see you here! |
samples/sample_03.wav |
Our meeting has been moved to three o'clock. |
samples/sample_04.wav |
She walked quietly across the empty room. |
samples/sample_05.wav |
How many people are coming to dinner tonight? |
24 kHz mono, 16-bit PCM.
Listen before trusting any metric. The word and character error rates in
EVAL.md measure whether an ASR model can recover the words. They say nothing
about naturalness, pacing or timbre. This model is a 3.76M-parameter vocoder
distilled from 3 hours of audio, and it sounds like it — the point is that it
is 4.8× smaller and 10× faster on an Arm CPU than the teacher, not that it
matches it.