From 2448c6f749670963c473656e339c63175a6662ba Mon Sep 17 00:00:00 2001
From: take0x <89313929+take0x@users.noreply.github.com>
Date: Mon, 9 Sep 2024 19:38:39 +0900
Subject: [PATCH 1/2] Transcribe on GPU

---
 whisper/transcribe.py | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/whisper/transcribe.py b/whisper/transcribe.py
index 1c075a201..d34152832 100644
--- a/whisper/transcribe.py
+++ b/whisper/transcribe.py
@@ -130,7 +130,9 @@ def transcribe(
         decode_options["fp16"] = False
 
     # Pad 30-seconds of silence to the input audio, for slicing
-    mel = log_mel_spectrogram(audio, model.dims.n_mels, padding=N_SAMPLES)
+    mel = log_mel_spectrogram(
+        audio, model.dims.n_mels, padding=N_SAMPLES, device=model.device
+    )
     content_frames = mel.shape[-1] - N_FRAMES
     content_duration = float(content_frames * HOP_LENGTH / SAMPLE_RATE)
 

From c1031a5787e7f21b789e9b84309d443d2fc7188a Mon Sep 17 00:00:00 2001
From: take0x <89313929+take0x@users.noreply.github.com>
Date: Mon, 23 Sep 2024 08:06:27 +0900
Subject: [PATCH 2/2] Add mel_spectrogram_device parameter

---
 whisper/transcribe.py | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/whisper/transcribe.py b/whisper/transcribe.py
index d34152832..d3a6283ab 100644
--- a/whisper/transcribe.py
+++ b/whisper/transcribe.py
@@ -51,6 +51,7 @@ def transcribe(
     append_punctuations: str = "\"'.。,，!！?？:：”)]}、",
     clip_timestamps: Union[str, List[float]] = "0",
     hallucination_silence_threshold: Optional[float] = None,
+    mel_spectrogram_device: Optional[Union[str, torch.device]] = None,
     **decode_options,
 ):
     """
@@ -113,6 +114,9 @@ def transcribe(
         When word_timestamps is True, skip silent periods longer than this threshold (in seconds)
         when a possible hallucination is detected
 
+    mel_spectrogram_device: Optional[Union[str, torch.device]]
+        If given, the audio tensor is moved to this device before STFT
+
     Returns
     -------
     A dictionary containing the resulting text ("text") and segment-level details ("segments"), and
@@ -131,7 +135,7 @@ def transcribe(
 
     # Pad 30-seconds of silence to the input audio, for slicing
     mel = log_mel_spectrogram(
-        audio, model.dims.n_mels, padding=N_SAMPLES, device=model.device
+        audio, model.dims.n_mels, padding=N_SAMPLES, device=mel_spectrogram_device
     )
     content_frames = mel.shape[-1] - N_FRAMES
     content_duration = float(content_frames * HOP_LENGTH / SAMPLE_RATE)