A hand-built speech-to-text engine, proven correct against the reference
He checked his rewrite against the original, layer by layer. In the middle of the network the numbers were off by two thousand. The transcript came out word for word identical.
One engine. 16+ model families. Your machine.
The project is transcribe dot c p p. You point it at an audio file. It writes out the text, on your own machine. One engine covers sixteen plus model families and sixty odd variants. One person wrote it.
He was shipping models he could not vouch for.
He also maintains Handy, a free dictation app with twenty-seven thousand stars. Shipping it meant handing users model files he found online. He could not tell whether they ran faithfully. So he checks every model this project publishes against its reference.
Same file. Two implementations. Eighteen checkpoints.
He ran one audio file through NVIDIA's code. He saved every tensor. Then he ran the same file through his rewrite. He compared the two at eighteen checkpoints. They do not track. His code computes the spectrogram in double precision. The reference uses single. He chose double on purpose, for this model's transform. The gap enters at the mel, at about five. The pre-encoder amplifies it to two thousand and eleven. Twelve blocks deep it is still around a thousand. Then by block twenty-three it is four hundredths. It is not smaller. It is gone from the encoder's output. This encoder is contractive. It squeezes the divergence back out. And the words that come out the other end are identical.
It lands on NVIDIA's own number.
One matching transcript does not prove it is accurate. So he measured it on the full test-clean split. It scores one point six eight percent word error. NVIDIA publishes one point six nine for the same model on the same split. It lands on the reference within rounding.
When the reference writes a transcript, only an exact match passes.
When the reference writes out its own transcript, the two must match exactly, character for character. And a missing reference file returns exit code seventy-seven, never green.
Slide 7
cmake -B build cmake --build build build/bin/transcribe-cli \ -m models/parakeet-tdt-0.6b-v2/parakeet-tdt-0.6b-v2-Q8_0.gguf \ samples/jfk.wav
It gives you the same answer.
It is version zero point one point three. The author says there are rough edges he cannot find alone. The part worth keeping is not the speech engine. It is that a correct rewrite does not give you the same numbers. It gives you the same answer.
