diff --git a/R/encoder.R b/R/encoder.R
index 284356a..258f687 100644
--- a/R/encoder.R
+++ b/R/encoder.R
@@ -86,8 +86,11 @@ whisper_attention <- torch::nn_module(
       attn_output <- torch::torch_matmul(attn_weights, v)
     } else {
       # Scaled dot-product attention (dispatches to FlashAttention on GPU)
-      attn_output <- torch:::torch_scaled_dot_product_attention(
-        q, k, v, is_causal = !is.null(mask))
+      # torch_scaled_dot_product_attention is not yet exported from torch
+      # (will be in next CRAN release). Use get() to avoid R CMD check NOTE.
+      sdpa <- get("torch_scaled_dot_product_attention",
+        envir = asNamespace("torch"))
+      attn_output <- sdpa(q, k, v, is_causal = !is.null(mask))
     }
 
     # Reshape back: (batch, n_head, seq_len, head_dim) -> (batch, seq_len, n_state)
diff --git a/cran-comments.md b/cran-comments.md
index ee63bf8..7081ec9 100644
--- a/cran-comments.md
+++ b/cran-comments.md
@@ -1,35 +1,33 @@
 ## R CMD check results
 
-0 errors | 0 warnings | 1 note
-
-* This is a new submission.
+0 errors | 0 warnings | 0 notes
 
 ## Test environments
 
-* local Ubuntu 24.04, R 4.4.x
+* local Ubuntu 24.04, R 4.5.2
 * GitHub Actions (r-ci): ubuntu-latest, macos-latest
+* win-builder R-devel
 
-## Package dependencies
-
-This package requires:
-- torch: Neural network operations
-- av: Audio file loading (FFmpeg bindings)
-- jsonlite: JSON parsing for tokenizer files
-- hfhub: HuggingFace model downloads
-- safetensors: Model weight loading
+## Changes since last CRAN release (0.1.0)
 
-## Resubmission changes
-
-- Added cornball.ai as copyright holder (`cph`) in `Authors@R` to match
-  the LICENSE file.
-- Added OpenAI as copyright holder (`cph`) in `Authors@R`. The bundled mel
-  filterbank data (`inst/assets/mel_80.csv`, `mel_128.csv`) is from OpenAI's
-  MIT-licensed Whisper repository, and the model architecture is derived from
-  their specifications.
-- Changed `\dontrun{}` to `\donttest{}` in `download_whisper_model()` examples.
+- `whisper_pipeline()` for cached model reuse across multiple transcriptions
+- SDPA attention (FlashAttention on GPU)
+- Segment-level and word-level timestamps via DTW alignment
+- Beam search decoding with temperature sampling and fallback
+- Automatic language detection (`detect_language()`, `language = NULL` default)
+- Hardcoded special token table (eliminates `added_tokens.json` download)
+- Fixed invalid multibyte string crash in BPE decoder
+- Fixed DTW boundary guards and seek loop in `transcribe_chunk()`
+- Fixed win-builder test crash: added `torch_is_installed()` guard
 
 ## Notes
 
+This package uses `torch_scaled_dot_product_attention()` from torch,
+which is not yet exported in the current CRAN release. We access it via
+`get("torch_scaled_dot_product_attention", envir = asNamespace("torch"))`.
+The function will be exported in the next torch release
+(PR: https://github.com/mlverse/torch/pull/1404).
+
 This package provides a native R implementation of OpenAI's Whisper
 speech-to-text model. Model weights are downloaded from HuggingFace
 on first use (ranging from 145MB for tiny to 3GB for large-v3).