whisper : fix sample_to_timestamp calculation with 64 bit precision to avoid overflow (#388)

boolemancer · ggerganov · web-flow · commit 08dc705a6942 · 2023-01-08T15:08:45.000+02:00
* Do calculation with 64 bit precision to avoid overflow

* Update whisper.cpp

Co-authored-by: Georgi Gerganov &lt;ggerganov@gmail.com&gt;

Co-authored-by: Georgi Gerganov &lt;ggerganov@gmail.com&gt;
diff --git a/whisper.cpp b/whisper.cpp
@@ -3388,7 +3388,7 @@ static int timestamp_to_sample(int64_t t, int n_samples) {
 }
 
 static int64_t sample_to_timestamp(int i_sample) {
-    return (100*i_sample)/WHISPER_SAMPLE_RATE;
+    return (100ll*i_sample)/WHISPER_SAMPLE_RATE;
 }
 
 // a cost-function / heuristic that is high for text that takes longer to pronounce

Original file line number	Diff line number	Diff line change
`@@ -3388,7 +3388,7 @@ static int timestamp_to_sample(int64_t t, int n_samples) {`
`3388`	`3388`	`}`
`3389`	`3389`
`3390`	`3390`	`static int64_t sample_to_timestamp(int i_sample) {`
`3391`		`- return (100*i_sample)/WHISPER_SAMPLE_RATE;`
	`3391`	`+ return (100ll*i_sample)/WHISPER_SAMPLE_RATE;`
`3392`	`3392`	`}`
`3393`	`3393`
`3394`	`3394`	`// a cost-function / heuristic that is high for text that takes longer to pronounce`