From 589abb9e465d2ca9eea5c019838d2435266ca3ac Mon Sep 17 00:00:00 2001
From: Michael Campagnaro <mikecampo@protonmail.com>
Date: Mon, 12 Jun 2023 22:49:45 -0400
Subject: [PATCH] Tweak whisper core/thread count

---
 bin/transcribe-audio | 16 ++++++++++++----
 bin/transcribe-video |  2 +-
 2 files changed, 13 insertions(+), 5 deletions(-)

diff --git a/bin/transcribe-audio b/bin/transcribe-audio
index f2be8c4..34ef892 100644
--- a/bin/transcribe-audio
+++ b/bin/transcribe-audio
@@ -71,16 +71,24 @@ fi
 
 output_name="$output_name_without_ext.${model}"
 
-printf "\n${YELLOW}${BOLD}Transcribing $input_wav | model: $model | threads: $threads | output: $output_name ${NORMAL}\n"
+
+# 1 core 31 threads has very fast pcm_to_mel conversion and then just one core doing most of the work. You get more accurate results this way.
+# 2 core 16 is about half the time but it can have errors where the two pieces come together. This only gets more likely as the core count is increased.
+# 8 threads, 4 cores is good too for tiny,small and 2 threads, 4 cores for medium.
+
+threads=31 #keep a thread for me
+cores=1
+
+printf "\n${YELLOW}${BOLD}Transcribing $input_wav | model: $model | cores: $cores | threads: $threads | output: $output_name ${NORMAL}\n"
 
 whisper_fullname="$(expand_path $(which whisper.exe))"
 whisper_path="$(dirname $whisper_fullname)"
 models_path="$whisper_path/models"
 
-whisper.exe --threads ${threads} -m "$models_path/ggml-${model}.en.bin" -otxt -osrt -f "$input_wav" -of "$output_name" --print-colors
+whisper.exe --processors ${cores} --threads ${threads} -m "$models_path/ggml-${model}.en.bin" -otxt -osrt -f "$input_wav" -of "$output_name" --print-colors --print-progress
 error=$?
 if [[ error -eq 0 ]]; then
-    printf "${GREEN}${BOLD}Done transcribing $input_wav | model: $model | threads: $threads | output: $output_name${NORMAL}\n"
+    printf "${GREEN}${BOLD}Done transcribing $input_wav | model: $model | cores: $cores | threads: $threads | output: $output_name${NORMAL}\n"
 else
-    printf "${GREEN}${BOLD}Error while transcribing $input_wav | model: $model | threads: $threads | output: $output_name${NORMAL}\n"
+    printf "${GREEN}${BOLD}Error while transcribing $input_wav | model: $model | cores: $cores | threads: $threads | output: $output_name${NORMAL}\n"
 fi
diff --git a/bin/transcribe-video b/bin/transcribe-video
index b300b10..871831f 100644
--- a/bin/transcribe-video
+++ b/bin/transcribe-video
@@ -29,7 +29,7 @@ shift 2
 models=("$@")
 
 if [[ $input_video == "" || $output_name_without_ext == "" || ${#models[@]} -eq 0 ]]; then
-    printf "${BOLD}${RED}Usage: $0 <input.wav> <output name without extension> <list of model names to use>${NORMAL}\n"
+    printf "${BOLD}${RED}Usage: $0 <input.mp4> <output name without extension> <list of model names to use>${NORMAL}\n"
     exit 1
 fi