From 8b0fb74bded1f8a162e6c0e94c3522be6216ea03 Mon Sep 17 00:00:00 2001
From: chengligen <101448376+chengligen@users.noreply.github.com>
Date: 星期一, 26 五月 2025 14:11:33 +0800
Subject: [PATCH] feat: add 'words' key aligned with timestamps in sensevoice model output (#2531)

---
 runtime/python/onnxruntime/funasr_onnx/sensevoice_bin.py |    6 ++++--
 1 files changed, 4 insertions(+), 2 deletions(-)

diff --git a/runtime/python/onnxruntime/funasr_onnx/sensevoice_bin.py b/runtime/python/onnxruntime/funasr_onnx/sensevoice_bin.py
index 9cfe46f..6a06ed1 100644
--- a/runtime/python/onnxruntime/funasr_onnx/sensevoice_bin.py
+++ b/runtime/python/onnxruntime/funasr_onnx/sensevoice_bin.py
@@ -183,8 +183,10 @@
                 #     ctc_logits = torch.from_numpy(ctc_logits).float()
                 # support batch_size=1 only currently
                 x = ctc_logits[b, : encoder_out_lens[b].item(), :]
-                yseq = x.argmax(dim=-1)
-                yseq = np.unique(yseq)
+                yseq = np.argmax(x, axis=-1)
+                # Use np.diff and np.where instead of torch.unique_consecutive.
+                mask = np.concatenate(([True], np.diff(yseq) != 0))
+                yseq = yseq[mask]
 
                 mask = yseq != self.blank_id
                 token_int = yseq[mask].tolist()

--
Gitblit v1.9.1