pyannote/segmentation-3.0 re-exported so that it can be loaded by tract, the pure-Rust ONNX runtime. Functionally identical to the upstream export; only the graph's shape handling differs.

Used by CallMind to measure how many speakers a recording holds instead of assuming two.

The published ONNX export cannot be loaded by tract. It contains an If node guarding a branch on the input's shape, plus symbolic dimension arithmetic that tract declines to prove equal (-24 + T/10 against -25 + (T+9)/10). Both disappear once the input is fixed to a single 10-second chunk: the branch condition becomes a constant and ordinary constant folding removes the node.

Verified against the source export:

Operators after folding are all standard ai.onnx: InstanceNormalization, Conv, MaxPool, LeakyRelu, Transpose, LSTM, Reshape, Gemm, LogSoftmax, Abs.

The int8 variant is deliberately not provided: it folds to a graph containing DynamicQuantizeLSTM from the com.microsoft domain, which tract does not implement.

The seven classes are ∅, {1}, {2}, {3}, {1,2}, {1,3}, {2,3} — up to three speakers with the two-at-once combinations, which is how overlapping speech is represented. Frame duration is 10000 / 589 ≈ 16.98 ms.

Longer audio is processed one chunk at a time; the final chunk is zero-padded.

Against labelled recordings — four single-speaker recordings confirmed by their owner and thirty two-party phone calls — taking the median number of distinct speakers seen per chunk:

The maximum is the wrong statistic despite looking natural: it is the maximum of a noisy quantity, so it grows with recording length and long calls reliably report one speaker too many.

The script re-downloads the source, applies the transformation and refuses to write a result whose output drifted, whose If node survived, or which contains non-standard operator domains.

MIT, inherited unchanged from upstream.

If you use pyannote in research, cite the upstream work: