hft-transformer-GGUF

Toyama et al.'s hierarchical frequency-time transformer for piano transcription (ISMIR 2023, sony/hFT-Transformer, MIT), converted to GGUF for the ggml runtime in CrispASR:

crispasr --piano -m hft-transformer-q4_0.gguf -f input.wav

Files

Quantisation is free here, q4_0 included. That is not the usual result and the reason is specific: the head q4_0 perturbs most is velocity (cosine 0.870), and in this model velocity feeds the decoder's ignore_zero gate rather than a note duration — so its errors change which notes survive, not how long they last. Contrast the Onsets & Frames GGUF, where q4_0 damages the frame head and costs 0.5 points of F1-with-offsets, because that head sets durations.

4.5 MiB reaching 70.7% on solo piano makes this the smallest model in this family that reaches that accuracy.

Accuracy and provenance

Converted from the pruned ONNX export of the sony/hFT-Transformer checkpoint. The converter fuses the (1, 5) convolution into tok_embedding_freq — both are linear in the 65-tap window with nothing between them, so they collapse exactly into one Linear(65, 256) (verified at 7.1e-07 relative).

The f32 build is the model, not an approximation of it: head-identical to the ONNX export (onset max abs 2.5e-06, cosine 1.00000000, 100.0000% of decisions identical including the velocity-gate argmax), and F1-identical on all ten pieces of MusicNet's test split.

Figures above are all ten MusicNet test pieces, 13,589 reference notes, scored with mir_eval.transcription (50 ms onset tolerance, 50 cents, maximum bipartite matching), every arm through the same decoder. Method and full tables: docs/music-transcription/HFT_TRANSFORMER.md and §35–36 of the CrispTuner report.

Cost — read this before choosing it

On a 4-vCPU Skylake-SP VPS, f32 is the fastest arm at 2.14× real time on four threads (4.84 CPU-s per audio-second single-threaded). Quantisation costs 29% of throughput while buying the size above, because ggml's repacked int8 GEMM is reached only through the CPU device's extra buffer types and this build does not request them — and that host has AVX-512F but no VNNI, so there is no int8 dot-product instruction to exploit either. The int8 case is untested here rather than disproved.

Native ONNX Runtime is 1.28× real time on the same box, so ggml is 1.56× its CPU cost — but at 237 MiB peak RSS against ORT's ~1001 MiB.

For throughput, onsets-and-frames is the better piano arm at 0.44× real time. Choose hFT for accuracy per megabyte, or where 4.5 MiB matters.