Why has whisper.cpp become less "fun" and useful for transcribing lyrics over music

[NielsMayer]

Whisper.cpp has been especially good at "getting" sometimes hard-to-understand patois derivatives of English in lyrics.

However, I recently updated my copy of whisper.cpp and got certain aspects of improvement in execution speed in tradeoff with the fact that I liked the output of the January 18 2023 version (aka :OLD) over the August 31, 2023 version (aka: NEW).

I've basically been using the "old" version (which I still have saved as .bak) to run Trainspodder transcriptions for rare cases where regular "faster-whisper" and RTX4080 doesn't do a good job, especially with such radio broadcasts with music, etc. See for example: https://rumble.com/v2nvr1w-trainspodder-helping-decode-reggae-lyrics.html ...

Are there specific default parameter settings that have changed, or command-line flags I can apply, to get the old behavior?

The commandline call producing the outputs below is essentially default:
whispercpp -t 8 -ocsv -m /opt/whisper/ggml-medium.en.bin -f <flewav>

OLD is ~ v1.4.2 January 18 2023: 206fc93
NEW from August 31, 2023: 2f52783

Note how much more useful, fun, descriptive (e.g. " (rapping in Spanish)" "(rapping in a foreign language)" "(upbeat music)") the OLD version is, compared with NEW. The transcription source material is from the BBC Radio1: https://www.bbc.co.uk/programmes/m001plth

OLD:

main: processing './Radio_1_s_Residency_BAMBII_-_Festival_Bangers_-_m001pltf.wav' (57616043 samples, 3601.0 sec), 8 threads, 1 processors, lang = en, task = transcribe, timestamps = 1 ...

[00:00:00.000 --> 00:00:20.000]   [MUSIC PLAYING]
[00:00:20.000 --> 00:00:23.000]   This program contains language that you might find offensive.
[00:00:23.000 --> 00:00:25.000]   ♪ You have been warned ♪
[00:00:25.000 --> 00:00:34.380]   ♪♪
[00:00:34.380 --> 00:00:35.880]   -Hello, everyone. I'm Bambi,
[00:00:35.880 --> 00:00:37.580]   and I'm so excited to be here with you
[00:00:37.580 --> 00:00:40.220]   for my residency on BBC Radio 1.
[00:00:40.220 --> 00:00:48.760]   ♪♪
[00:00:48.760 --> 00:00:51.360]   Last few episodes, I've taken you through
[00:00:51.360 --> 00:00:56.900]   What's influenced me, the producers I'm listening to, and this theme is festival.
[00:00:56.900 --> 00:01:03.260]   I just played two amazing shows at Pucko Pop and Lolan, so I decided to theme the whole mix around
[00:01:03.260 --> 00:01:06.060]   mainstage driving tracks.
[00:01:06.060 --> 00:01:14.800]   It's also unfortunately the last episode of my residency here at BBC Radio 1. I want to go out with a bang.
[00:01:14.800 --> 00:01:41.760]   [MUSIC PLAYING]
[00:01:41.760 --> 00:01:45.760]   This is BBC Radio 1's residency.
[00:01:45.760 --> 00:01:48.340]   (upbeat music)
[00:01:48.340 --> 00:01:50.920]   (upbeat music)
[00:01:50.920 --> 00:01:53.500]   (upbeat music)
[00:01:53.500 --> 00:01:56.080]   (upbeat music)
[00:01:56.080 --> 00:01:58.660]   (upbeat music)
[00:01:58.660 --> 00:02:01.240]   (upbeat music)
[00:02:01.240 --> 00:02:03.820]   (upbeat music)
[00:02:03.820 --> 00:02:06.400]   (upbeat music)
[00:02:06.400 --> 00:02:08.980]   (upbeat music)
[00:02:08.980 --> 00:02:11.560]   (upbeat music)
[00:02:11.560 --> 00:02:14.140]   (upbeat music)
[00:02:14.140 --> 00:02:16.720]   (upbeat music)
[00:02:16.720 --> 00:02:19.300]   (upbeat music)
[00:02:45.460 --> 00:02:47.460]   (music)
[00:02:47.460 --> 00:03:17.460]   T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T-T
[00:03:17.460 --> 00:03:27.460]   ♪♪
[00:03:27.460 --> 00:03:37.480]   ♪♪
[00:03:37.480 --> 00:03:40.280]   He always tell me follow me, he gon' swallow me
[00:03:40.280 --> 00:03:41.880]   Yo, she want Christian Dior
[00:03:41.880 --> 00:03:43.880]   I took me with you, yeah, I underscore
[00:03:43.880 --> 00:03:45.480]   Fuck 'em good, head 'em begging for more
[00:03:45.480 --> 00:03:47.480]   Then I tip, tip, tip, tip, then I tip through the door
[00:03:47.480 --> 00:03:49.480]   Yo, tip, tip through the door
[00:03:49.480 --> 00:03:51.480]   Stirring up core
[00:03:51.480 --> 00:03:53.480]   Tip, tip, tip through the door
[00:03:53.480 --> 00:03:55.480]   Stirring up core
[00:03:55.480 --> 00:03:57.480]   Tip, tip through the door
[00:03:57.480 --> 00:03:59.480]   Stirring up core
[00:03:59.480 --> 00:04:01.480]   Tip, tip, tip through the door
[00:04:01.480 --> 00:04:03.480]   Tip, tip, tip through the door
[00:04:03.480 --> 00:04:05.480]   Stirring up raw
[00:04:05.480 --> 00:04:09.480]   Tip, tip, tip through the door, serving up raw
[00:04:09.480 --> 00:04:12.480]   Tip, tip, tip through the door, serving up raw
[00:04:12.480 --> 00:04:15.480]   Tip, tip, tip through the door
[00:04:15.480 --> 00:04:18.480]   Me and Vibe, rob in the face of this shit
[00:04:18.480 --> 00:04:20.480]   We coming out these hoes, what's up they bitch?
[00:04:20.480 --> 00:04:22.480]   Eyes on the prize, put your hands on my hip
[00:04:22.480 --> 00:04:24.480]   You ain't no Jersey boy, get it rocking like this
[00:04:24.480 --> 00:04:25.480]   Hit a gritful, I lose my patience
[00:04:25.480 --> 00:04:27.480]   I pack the line, I'm so far from basic
[00:04:27.480 --> 00:04:29.480]   Body insane and it look vivacious
[00:04:29.480 --> 00:04:31.480]   Now you all up in my ear, speaking real salacious
[00:04:31.480 --> 00:04:33.480]   Name another bad, bad one like me
[00:04:33.480 --> 00:04:34.480]   Tip through the door, now they let out the beast
[00:04:34.480 --> 00:04:36.480]   See me on the floor getting down with the freaks
[00:04:36.480 --> 00:04:38.480]   And the pistol on leak, housewife with the peach
[00:04:38.480 --> 00:04:40.480]   VIP giving high class tees
[00:04:40.480 --> 00:04:42.480]   Pour my shot while I smoke my trees
[00:04:42.480 --> 00:04:45.480]   Off IG now I'm in HD so wishfully I get busy like me
[00:04:45.480 --> 00:04:47.480]   Last to arrive and the first one to leave
[00:04:47.480 --> 00:04:49.480]   Rats got the bite but they don't grab the cheese
[00:04:49.480 --> 00:04:51.480]   From the underground to your radio station
[00:04:51.480 --> 00:04:53.480]   Blowing up a tab I need new vibrations
[00:04:53.480 --> 00:04:55.480]   Sock you should see me serve it up more
[00:04:55.480 --> 00:04:57.480]   Style got me draped like a real label war
[00:04:57.480 --> 00:04:59.480]   Take this pic while I sip my pour
[00:04:59.480 --> 00:05:01.480]   Hee-kee-kee-kee-kip give me some more
[00:05:01.480 --> 00:05:03.480]   Tip tip tip you a door
[00:05:03.480 --> 00:05:07.120]   ♪ Serving up for her, tip, tip, tip through the door ♪
[00:05:07.120 --> 00:05:10.820]   ♪ Serving up for her, tip, tip, tip through the door ♪
[00:05:10.820 --> 00:05:13.900]   ♪ Serving up for her ♪
[00:05:13.900 --> 00:05:28.740]   - That last one was an amazing unreleased track
[00:05:28.740 --> 00:05:31.020]   by Bianca of Livian called "Tales of Glory."
[00:05:31.020 --> 00:05:32.780]   I've said it before, I'll say it again.
[00:05:32.780 --> 00:05:35.540]   I never DJ without playing a Bianca oblivion track.
[00:05:35.540 --> 00:05:38.540]   She's honestly, to me, one of the best club producers
[00:05:38.540 --> 00:05:40.060]   out right now.
[00:05:40.060 --> 00:05:42.900]   If you don't know her, you definitely have to check her out.

NEW:

system_info: n_threads = 8 / 32 | AVX = 1 | AVX2 = 1 | AVX512 = 0 | FMA = 1 | NEON = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0| BLAS = 0 | SSE3 = 1 | SSSE3 = 1 | VSX = 0 | COREML = 0 | OPENVINO = 0 |

main: processing 'Radio_1_s_Residency_BAMBII_-_Festival_Bangers_-_m001pltf.wav' (57616043 samples, 3601.0 sec), 8 threads, 1 processors, lang= en, task = transcribe, timestamps = 1 ...


[00:00:00.000 --> 00:00:20.000]   [MUSIC PLAYING]
[00:00:20.000 --> 00:00:23.000]   This program contains language that you might find offensive.
[00:00:23.000 --> 00:00:34.000]   You have been warned.
[00:00:34.000 --> 00:00:35.000]   Hello, everyone.
[00:00:35.000 --> 00:00:40.000]   I'm Bambi, and I'm so excited to be here with you from my residency on BBC Radio 1.
[00:00:40.000 --> 00:00:48.000]   [MUSIC PLAYING]
[00:00:48.000 --> 00:00:52.000]   Last few episodes, I've taken you through what's influenced me,
[00:00:52.000 --> 00:00:57.000]   the producers I'm listening to, and this theme is festival.
[00:00:57.000 --> 00:01:01.000]   I just played two amazing shows at Puckle Pop and Lolan,
[00:01:01.000 --> 00:01:08.000]   so I decided to theme the whole mix around main stage driving tracks.
[00:01:08.000 --> 00:01:13.000]   It's also unfortunately the last episode of my residency here at BBC Radio 1.
[00:01:13.000 --> 00:01:15.000]   I want to go out with a bang.
[00:01:15.000 --> 00:01:42.000]   [MUSIC PLAYING]
[00:01:42.000 --> 00:01:46.000]   This is BBC Radio 1's residency.

[00:01:47.000 --> 00:02:15.000]   [MUSIC PLAYING]
[00:02:15.000 --> 00:02:44.000]   [MUSIC PLAYING]
[00:02:44.000 --> 00:02:55.000]   [MUSIC PLAYING]
[00:02:55.000 --> 00:03:09.000]   [MUSIC PLAYING]
[00:03:09.000 --> 00:03:20.000]   [MUSIC PLAYING]
[00:03:20.000 --> 00:03:30.000]   [MUSIC PLAYING]
[00:03:30.000 --> 00:03:41.000]   [MUSIC PLAYING]
[00:03:41.000 --> 00:03:51.000]   [MUSIC PLAYING]
[00:03:51.000 --> 00:04:02.000]   [MUSIC PLAYING]
[00:04:02.000 --> 00:04:12.000]   [MUSIC PLAYING]
[00:04:12.000 --> 00:04:19.000]   [MUSIC PLAYING]
[00:04:19.000 --> 00:04:29.000]   [MUSIC PLAYING]
[00:04:29.000 --> 00:04:39.000]   [MUSIC PLAYING]
[00:04:39.000 --> 00:04:43.000]   [MUSIC PLAYING]
[00:04:43.000 --> 00:04:53.000]   [MUSIC PLAYING]
[00:04:53.000 --> 00:05:03.000]   [MUSIC PLAYING]
[00:05:03.000 --> 00:05:13.000]   [MUSIC PLAYING]
[00:05:13.000 --> 00:05:23.000]   [MUSIC PLAYING]
[00:05:23.000 --> 00:05:26.000]   [MUSIC PLAYING]
[00:05:26.000 --> 00:05:31.000]   That last one was an amazing unreleased track by Bianca of Livia called Tales of Glory.
[00:05:31.000 --> 00:05:32.000]   I've said it before, I'll say it again.
[00:05:32.000 --> 00:05:35.000]   I never DJ without playing a Beyonce oblivion track.
[00:05:35.000 --> 00:05:40.000]   She's honestly, to me, one of the best club producers out right now.
[00:05:40.000 --> 00:05:42.000]   If you don't know her, you definitely have to check her out.

Another comparison of a different section:

OLD:

[00:27:06.440 --> 00:27:36.440]   Take the one biscuit, take the one biscuit, take the one biscuit, take the one biscuit, take the one biscuit, take the one biscuit, take the one biscuit, take the one biscuit, take the one biscuit, take the one biscuit, take the one biscuit, take the one biscuit, take the one biscuit, take the one biscuit, take the one biscuit, take the one biscuit, take the one biscuit, take the one biscuit, take the one biscuit, take the one biscuit, take the one biscuit, take the one biscuit, take the one biscuit, take the one biscuit, take the one biscuit, take the one biscuit, take the one biscuit, take the one biscuit, take the one biscuit, take the one biscuit, take the one biscuit, take the one biscuit, take the one biscuit, take the one biscuit, take the one biscuit, take the one biscuit, take the one
[00:27:36.440 --> 00:27:38.940]   (beat-boxing)
[00:27:38.940 --> 00:28:08.940]   Fake the Lemmy Skit, Fake the Lemmy Skit, Fake the Lemmy Skit, Fake the Lemmy Skit, Fake the Lemmy Skit, Fake the Lemmy Skit, Fake the Lemmy Skit, Fake the Lemmy Skit, Fake the Lemmy Skit, Fake the Lemmy Skit, Fake the Lemmy Skit, Fake the Lemmy Skit, Fake the Lemmy Skit, Fake the Lemmy Skit, Fake the Lemmy Skit, Fake the Lemmy Skit, Fake the Lemmy Skit, Fake the Lemmy Skit, Fake the Lemmy Skit, Fake the Lemmy Skit, Fake the Lemmy Skit, Fake the Lemmy Skit, Fake the Lemmy Skit, Fake the Lemmy Skit, Fake the Lemmy Skit, Fake the Lemmy Skit, Fake the Lemmy Skit, Fake the Lemmy Skit, Fake the Lemmy Skit, Fake the Lemmy Skit, Fake the Lemmy Skit, Fake the
[00:28:08.940 --> 00:28:38.940]   Get the ones, get the ones, get the ones, get the ones, get the ones, get the ones, get the ones, get the ones, get the ones, get the ones, get the ones, get the ones, get the ones, get the ones, get the ones, get the ones, get the ones, get the ones, get the ones, get the ones, get the ones, get the ones, get the ones, get the ones, get the ones, get the ones, get the ones, get the ones, get the ones, get the ones, get the ones, get the ones, get the ones, get the ones, get the ones, get the ones, get the ones, get the ones, get the ones, get the ones, get the ones, get the ones, get the ones, get the ones, get the ones, get the ones, get the ones, get the ones,get the ones, get the ones, get the ones, get the ones, get the ones, get the ones, get the ones
[00:28:38.940 --> 00:28:40.940]   I'm the one
[00:28:40.940 --> 00:28:44.940]   I'm the only one
[00:28:44.940 --> 00:28:57.940]   I'm the...
[00:28:57.940 --> 00:29:00.940]   This is BBC Radio 1's residency
[00:29:00.940 --> 00:29:03.940]   Do you need a kick?
[00:29:03.940 --> 00:29:06.940]   Do you need a kick?
[00:29:06.940 --> 00:29:08.940]   (In the neck)
[00:29:08.940 --> 00:29:10.940]   (In the neck)
[00:29:10.940 --> 00:29:12.940]   (In the neck)
[00:29:12.940 --> 00:29:14.940]   (In the neck)
[00:29:14.940 --> 00:29:16.940]   (In the neck)
[00:29:16.940 --> 00:29:18.940]   Too classy for this world
[00:29:18.940 --> 00:29:20.940]   Too gassed up on that girl
[00:29:20.940 --> 00:29:22.940]   Teaching time and time clouds
[00:29:22.940 --> 00:29:24.940]   Ooh baby, ah
[00:29:24.940 --> 00:29:26.940]   Too classy to be touched
[00:29:26.940 --> 00:29:28.940]   I think I'm all in dust
[00:29:28.940 --> 00:29:30.940]   I'm still here with my love
[00:29:30.940 --> 00:29:32.940]   On a plane
[00:29:32.940 --> 00:29:43.940]   Unique
[00:29:43.940 --> 00:29:47.940]   Unique
[00:29:47.940 --> 00:29:53.940]   Unique
[00:29:53.940 --> 00:29:57.940]   Unique
[00:29:57.940 --> 00:30:04.940]   You made it, you made it You made it, you made it
[00:30:04.940 --> 00:30:12.940]   You made it, you made it You made it, you made it
[00:30:12.940 --> 00:30:15.520]   (upbeat music)
[00:30:15.520 --> 00:30:18.100]   (upbeat music)
[00:30:18.100 --> 00:30:21.100]   I wanna, I wanna, I wanna, I wanna
[00:30:21.100 --> 00:30:29.100]   Pretty little bitch, gimme money, gimme all that
[00:30:29.100 --> 00:30:31.100]   Pretty little bitch, gimme money, gimme all that
[00:30:31.100 --> 00:30:33.100]   Pretty little bitch, gimme money, gimme all that
[00:30:33.100 --> 00:30:35.100]   Pretty little bitch, little bitch, gimme money, gimme all that
[00:30:35.100 --> 00:30:37.100]   Pretty little bitch, gimme money, gimme all that
[00:30:37.100 --> 00:30:39.100]   Pretty little bitch, gimme money, gimme all that
[00:30:39.100 --> 00:30:40.100]   Pretty little bitch, gimme money, gimme all that
[00:30:40.100 --> 00:30:42.100]   Pretty little bitch, gimme money, gimme money, gimme all that
[00:30:42.100 --> 00:30:44.100]   I wanna, I wanna, I wanna, I wanna, I wanna
[00:30:44.100 --> 00:30:46.100]   I wanna, I wanna, I wanna, I wanna
[00:30:46.100 --> 00:30:50.100]   I wanna run us with the picky homie I wanna run us with the picky homie
[00:30:50.100 --> 00:30:53.100]   I wanna run us with the picky homie Tell me what you want
[00:30:53.100 --> 00:30:55.100]   Tell me what you want I wanna run us with the picky homie
[00:30:55.100 --> 00:30:58.100]   I wanna run us with the pretty little bitch gimme pretty gimme
[00:30:58.100 --> 00:31:00.100]   Pretty little bitch gimme pretty gimme pretty gimme pretty
[00:31:00.100 --> 00:31:02.100]   Pretty little bitch gimme pretty gimme pretty gimme
[00:31:02.100 --> 00:31:04.100]   Little bitch little bitch little bitch little bitch
[00:31:04.100 --> 00:31:06.100]   Pretty little bitch gimme pretty gimme pretty
[00:31:06.100 --> 00:31:08.100]   Little bitch gimme pretty gimme pretty little bitch gimme
[00:31:08.100 --> 00:31:11.100]   Pretty little bitch gimme pretty gimme pretty gimme pretty pretty
[00:31:11.100 --> 00:31:15.100]   I wanna run us with the picky homie I wanna run us with the picky homie
[00:31:15.100 --> 00:31:17.100]   ♪ I want a manos with the peque on me ♪
[00:31:17.100 --> 00:31:19.100]   ♪ I want a manos with the peque on me ♪
[00:31:19.100 --> 00:31:21.100]   ♪ I want a manos with the peque on me ♪
[00:31:21.100 --> 00:31:22.100]   ♪ With the peque on me ♪
[00:31:22.100 --> 00:31:23.100]   ♪ With the peque on me ♪
[00:31:23.100 --> 00:31:24.100]   ♪ I want a manos with the peque on me ♪
[00:31:24.100 --> 00:31:26.100]   ♪ I want a manos with the peque on me ♪
[00:31:26.100 --> 00:31:31.100]   You are listening to A Wasted Eyes remix by Ace Manual.
[00:31:31.100 --> 00:31:34.100]   Definitely an underground club producer staple.
[00:31:34.100 --> 00:31:36.100]   I feel like everyone plays his music.
[00:31:36.100 --> 00:31:39.100]   I had the pleasure of DJing with him years ago.
[00:31:39.100 --> 00:31:41.100]   He's an amazing DJ, amazing producer.
[00:31:41.100 --> 00:31:42.100]   Definitely check him out.
[00:31:42.100 --> 00:31:44.100]   ♪ Under me, I know you see him on me ♪
[00:31:44.100 --> 00:31:46.100]   I think of you, you're the real me
[00:31:46.100 --> 00:31:48.100]   I'm a formidable, make a best of
[00:31:48.100 --> 00:31:50.100]   Baby, you're the real me, you see my scene
[00:31:50.100 --> 00:31:52.100]   Under the radar, you scream on me
[00:31:52.100 --> 00:31:54.100]   I think of you, you're the real me
[00:31:54.100 --> 00:31:55.100]   Sorry, I messed up, tell me what you want
[00:31:55.100 --> 00:31:56.100]   She want more than I say
[00:31:56.100 --> 00:31:57.100]   I think of you, you're the real me
[00:31:57.100 --> 00:31:59.100]   Pretty little bitch, give me plenty, give me money
[00:31:59.100 --> 00:32:01.100]   Take me to no one, I'm just a little high
[00:32:01.100 --> 00:32:03.100]   I can't be her love
[00:32:03.100 --> 00:32:04.100]   So many things for you
[00:32:04.100 --> 00:32:06.100]   You love me more
[00:32:06.100 --> 00:32:08.100]   But I wanna leave my home
[00:32:08.100 --> 00:32:10.680]   (upbeat music)
[00:32:10.680 --> 00:32:13.260]   (upbeat music)
[00:32:13.260 --> 00:32:15.840]   (upbeat music)
[00:32:15.840 --> 00:32:18.420]   (upbeat music)
[00:32:18.420 --> 00:32:21.000]   (upbeat music)
[00:32:21.000 --> 00:32:23.580]   (upbeat music)
[00:32:23.580 --> 00:32:26.660]   (rapping in Spanish)
[00:32:26.660 --> 00:32:29.240]   (upbeat music)
[00:32:29.240 --> 00:32:31.820]   (upbeat music)
[00:32:31.820 --> 00:32:34.400]   (upbeat music)
[00:32:34.400 --> 00:32:36.980]   (upbeat music)

[00:32:36.980 --> 00:32:38.740]   We're excited, we're drunk
[00:32:38.740 --> 00:32:40.420]   We're gonna live, we're drunk
[00:32:40.420 --> 00:32:42.220]   We're gonna have fun, we're drunk
[00:32:42.220 --> 00:32:44.800]   (upbeat music)
[00:32:44.800 --> 00:32:47.380]   (upbeat music)
[00:32:47.380 --> 00:32:49.960]   (upbeat music)
[00:32:49.960 --> 00:32:52.540]   (upbeat music)
[00:32:52.540 --> 00:32:55.120]   (upbeat music)
[00:32:55.120 --> 00:33:05.120]   ♪♪
[00:33:05.120 --> 00:33:15.140]   ♪♪
[00:33:15.140 --> 00:33:17.720]   (upbeat music)
[00:33:17.720 --> 00:33:20.300]   (upbeat music)
[00:33:46.020 --> 00:33:48.780]   This is BBC Radio 1's Residence.
[00:33:48.780 --> 00:33:51.360]   (upbeat music)
[00:33:51.360 --> 00:33:57.480]   Wait a second, it's been a minute Got the first six, what's the last four digits?
[00:33:57.480 --> 00:33:57.980]   (Hey!)
[00:33:57.980 --> 00:34:03.000]   10 killers, doctors said feel better Picking up our habits, let's go
[00:34:03.000 --> 00:34:09.360]   Fuck the system, what's the rhythm? What's the rhythm? What's the rhythm?
[00:34:09.360 --> 00:34:11.800]   What's the, what's the, what's the rhythm?
[00:34:11.800 --> 00:34:13.800]   ♪♪
[00:34:13.800 --> 00:34:15.800]   ♪♪
[00:34:15.800 --> 00:34:17.800]   ♪♪
[00:34:17.800 --> 00:34:19.800]   ♪♪
[00:34:19.800 --> 00:34:21.800]   ♪♪
[00:34:21.800 --> 00:34:23.800]   ♪♪
[00:34:23.800 --> 00:34:25.800]   ♪♪
[00:34:25.800 --> 00:34:27.800]   ♪♪
[00:34:27.800 --> 00:34:29.800]   ♪♪
[00:34:29.800 --> 00:34:31.800]   ♪♪
[00:34:31.800 --> 00:34:33.800]   ♪♪
[00:34:33.800 --> 00:34:35.800]   ♪♪
[00:34:35.800 --> 00:34:37.800]   ♪♪
[00:34:37.800 --> 00:34:39.800]   ♪♪
[00:34:39.800 --> 00:34:44.300]   ♪ No, no, no, no, no, no, no, no ♪
[00:34:44.300 --> 00:34:47.800]   ♪ Monkeys coming round round round round round round ♪
[00:34:47.800 --> 00:34:52.300]   ♪ Shoutin' out to grandpa, come on, come on, come on ♪
[00:34:52.300 --> 00:34:55.300]   ♪ Sweetest second, I've been missin' ♪
[00:34:55.300 --> 00:34:59.300]   ♪ Bitch, I'm comin' out like I'm not much pregnant ♪
[00:34:59.300 --> 00:35:02.300]   ♪ Playing tickets like a grandeur wishes ♪
[00:35:02.300 --> 00:35:05.800]   ♪ Drivin' off the clique 'cause she givin' me the signals ♪
[00:35:05.800 --> 00:35:08.300]   ♪ This is the phone call, I made schoolte, what do you say? ♪
[00:35:08.300 --> 00:35:10.880]   (upbeat music)
[00:35:10.880 --> 00:35:12.880]   (music)

NEW -- note it's completely missing stuff that the old version actually did a great job with!

[00:27:09.000 --> 00:27:19.000]   [MUSIC PLAYING]
[00:27:19.000 --> 00:27:29.000]   [MUSIC PLAYING]
[00:27:29.000 --> 00:27:39.000]   [MUSIC PLAYING]
[00:27:39.000 --> 00:27:49.000]   [MUSIC PLAYING]
[00:27:49.000 --> 00:27:59.000]   [MUSIC PLAYING]
[00:27:59.000 --> 00:28:12.600]   (beatboxing)
[00:28:12.600 --> 00:28:26.060]   (beatboxing)
[00:28:26.060 --> 00:28:36.060]   (beatboxing)
[00:28:36.060 --> 00:28:46.060]   (beatboxing)
[00:28:46.060 --> 00:28:56.060]   (beatboxing)
[00:28:56.060 --> 00:29:06.060]   (beatboxing)
[00:29:06.060 --> 00:29:16.060]   (beatboxing)
[00:29:16.060 --> 00:29:26.060]   (beatboxing)
[00:29:26.060 --> 00:29:36.060]   (beatboxing)
[00:29:36.060 --> 00:29:46.060]   (beatboxing)
[00:29:46.060 --> 00:29:56.060]   (beatboxing)
[00:29:56.060 --> 00:30:06.060]   (beatboxing)
[00:30:06.060 --> 00:30:16.060]   (beatboxing)
[00:30:16.060 --> 00:30:26.060]   (beatboxing)
[00:30:26.060 --> 00:30:36.060]   (beatboxing)
[00:30:36.060 --> 00:30:46.060]   (beatboxing)
[00:30:46.060 --> 00:30:56.060]   (beatboxing)
[00:30:56.060 --> 00:31:06.060]   (beatboxing)
[00:31:06.060 --> 00:31:16.060]   (beatboxing)
[00:31:16.060 --> 00:31:26.060]   (beatboxing)
[00:31:26.060 --> 00:31:31.060]   You are listening to a Wasted Eyes remix by Ace Manual.
[00:31:31.060 --> 00:31:35.060]   Definitely an underground club producer staple.
[00:31:35.060 --> 00:31:39.060]   Everyone plays his music. I had the pleasure of DJing with him years ago.
[00:31:39.060 --> 00:31:43.060]   He's an amazing DJ, amazing producer. Definitely check him out.
[00:31:43.060 --> 00:31:53.060]   (beatboxing)
[00:31:53.060 --> 00:32:03.060]   (beatboxing)
[00:32:03.060 --> 00:32:13.060]   (beatboxing)
[00:32:13.060 --> 00:32:23.060]   (beatboxing)
[00:32:23.060 --> 00:32:33.060]   (beatboxing)
[00:32:33.060 --> 00:32:43.060]   (beatboxing)
[00:32:43.060 --> 00:32:53.060]   (beatboxing)
[00:32:53.060 --> 00:33:03.060]   (beatboxing)
[00:33:03.060 --> 00:33:13.060]   (beatboxing)
[00:33:13.060 --> 00:33:23.060]   (beatboxing)
[00:33:23.060 --> 00:33:33.060]   (beatboxing)
[00:33:33.060 --> 00:33:43.060]   (beatboxing)
[00:33:43.060 --> 00:33:47.060]   (beatboxing)
[00:33:47.060 --> 00:33:57.060]   (beatboxing)
[00:33:57.060 --> 00:34:07.060]   (beatboxing)
[00:34:07.060 --> 00:34:17.060]   (beatboxing)
[00:34:17.060 --> 00:34:27.060]   (beatboxing)
[00:34:27.060 --> 00:34:37.060]   (beatboxing)
[00:34:37.060 --> 00:34:47.060]   (beatboxing)
[00:34:47.060 --> 00:34:57.060]   (beatboxing)
[00:34:57.060 --> 00:35:07.060]   (beatboxing)
[00:35:07.060 --> 00:35:17.060]   (beatboxing)

[bobqianic]

Great catch! I think there are three likely explanations for this issue: it could be the change in #473, the one in #425, or the one in #1148. Personally, I'm leaning towards the first two options because they seem to have a more direct impact on the transcription text output. My change in #1148 mainly ensures that the output of the log_mel_spectrogram aligns with OpenAI's standards.

https://github.com/ggerganov/whisper.cpp/blob/2f52783a080e8955e80e4324fed73e2f906bb80c/whisper.cpp#L3747-L3766

https://github.com/ggerganov/whisper.cpp/blob/2f52783a080e8955e80e4324fed73e2f906bb80c/whisper.cpp#L3800-L3809

[bobqianic]

Have there been any updates? If something has broken this useful feature, we can easily implement a fallback option.

[bobqianic]

I will conduct some further testing today. Let's see what causes this issue. @NielsMayer

[bobqianic]

Okay, I've figured out what's causing this issue! I've successfully replicated the behavior of v1.4.2 in the latest master version. I'll try to explain the reason later. @NielsMayer

Before Fix: ggml-model-whisper-base.bin

C:\Users\qianp\Downloads\whisper.cpp_build-fix\bin\Release>C:\Users\qianp\Downloads\whisper.cpp_build-fix\bin\Release\main.exe -m C:\Users\qianp\Downloads\whisper.cpp_build-fix\bin\Release\ggml-model-whisper-base.bin -f "C:\Users\qianp\Downloads\whisper.cpp_build-fix\bin\Release\Coldplay - Viva La Vida (Official Video).wav" -t 4 -bs 5 -bo 5
whisper_init_from_file_no_state: loading model from 'C:\Users\qianp\Downloads\whisper.cpp_build-fix\bin\Release\ggml-model-whisper-base.bin'
whisper_model_load: loading model
whisper_model_load: n_vocab       = 51865
whisper_model_load: n_audio_ctx   = 1500
whisper_model_load: n_audio_state = 512
whisper_model_load: n_audio_head  = 8
whisper_model_load: n_audio_layer = 6
whisper_model_load: n_text_ctx    = 448
whisper_model_load: n_text_state  = 512
whisper_model_load: n_text_head   = 8
whisper_model_load: n_text_layer  = 6
whisper_model_load: n_mels        = 80
whisper_model_load: ftype         = 1
whisper_model_load: qntvr         = 0
whisper_model_load: type          = 2
whisper_model_load: mem required  =  310.00 MB (+    6.00 MB per decoder)
whisper_model_load: adding 1608 extra tokens
whisper_model_load: model ctx     =  140.66 MB
whisper_model_load: model size    =  140.54 MB
whisper_init_state: kv self size  =    5.25 MB
whisper_init_state: kv cross size =   17.58 MB

system_info: n_threads = 4 / 20 | AVX = 1 | AVX2 = 1 | AVX512 = 0 | FMA = 1 | NEON = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 1 | SSSE3 = 0 | VSX = 0 | COREML = 0 | OPENVINO = 0 |

main: processing 'C:\Users\qianp\Downloads\whisper.cpp_build-fix\bin\Release\Coldplay - Viva La Vida (Official Video).wav' (3873646 samples, 242.1 sec), 4 threads, 1 processors, lang = en, task = transcribe, timestamps = 1 ...


[00:00:00.000 --> 00:00:10.000]   [MUSIC]
[00:00:10.000 --> 00:00:20.000]   [MUSIC]
[00:00:20.000 --> 00:00:30.000]   [MUSIC]
[00:00:30.000 --> 00:00:38.000]   [MUSIC]
[00:00:38.000 --> 00:00:48.000]   [MUSIC]
[00:00:48.000 --> 00:00:58.000]   [MUSIC]
[00:00:58.000 --> 00:01:08.000]   [MUSIC]
[00:01:08.000 --> 00:01:18.000]   [MUSIC]
[00:01:18.000 --> 00:01:28.000]   [MUSIC]
[00:01:28.000 --> 00:01:38.000]   [MUSIC]
[00:01:38.000 --> 00:01:48.000]   [MUSIC]
[00:01:48.000 --> 00:01:58.000]   [MUSIC]
[00:01:58.000 --> 00:02:08.000]   [MUSIC]
[00:02:08.000 --> 00:02:18.000]   [MUSIC]
[00:02:18.000 --> 00:02:28.000]   [MUSIC]
[00:02:28.000 --> 00:02:38.000]   [MUSIC]
[00:02:38.000 --> 00:02:48.000]   [MUSIC]
[00:02:48.000 --> 00:02:58.000]   [MUSIC]
[00:02:58.000 --> 00:03:08.000]   [MUSIC]
[00:03:08.000 --> 00:03:18.000]   [MUSIC]
[00:03:18.000 --> 00:03:28.000]   [MUSIC]
[00:03:28.000 --> 00:03:38.000]   [MUSIC]
[00:03:38.000 --> 00:03:48.000]   [MUSIC]
[00:03:48.000 --> 00:03:58.000]   [MUSIC]
[00:03:58.000 --> 00:04:00.580]   (gentle music)
[00:04:00.580 --> 00:04:10.580]   [MUSIC]


whisper_print_timings:     load time =    82.92 ms
whisper_print_timings:     fallbacks =   0 p /   0 h
whisper_print_timings:      mel time =   146.07 ms
whisper_print_timings:   sample time =   719.42 ms /  1010 runs (    0.71 ms per run)
whisper_print_timings:   encode time = 18978.69 ms /    24 runs (  790.78 ms per run)
whisper_print_timings:   decode time =  5095.26 ms /   906 runs (    5.62 ms per run)
whisper_print_timings:    total time = 25076.54 ms

After Fix: ggml-model-whisper-base.bin

C:\Users\qianp\Downloads\whisper.cpp_build-fix\bin\Release>C:\Users\qianp\Downloads\whisper.cpp_build-fix\bin\Release\main.exe -m C:\Users\qianp\Downloads\whisper.cpp_build-fix\bin\Release\ggml-model-whisper-base.bin -f "C:\Users\qianp\Downloads\whisper.cpp_build-fix\bin\Release\Coldplay - Viva La Vida (Official Video).wav" -t 4 -bs 5 -bo 5
whisper_init_from_file_no_state: loading model from 'C:\Users\qianp\Downloads\whisper.cpp_build-fix\bin\Release\ggml-model-whisper-base.bin'
whisper_model_load: loading model
whisper_model_load: n_vocab       = 51865
whisper_model_load: n_audio_ctx   = 1500
whisper_model_load: n_audio_state = 512
whisper_model_load: n_audio_head  = 8
whisper_model_load: n_audio_layer = 6
whisper_model_load: n_text_ctx    = 448
whisper_model_load: n_text_state  = 512
whisper_model_load: n_text_head   = 8
whisper_model_load: n_text_layer  = 6
whisper_model_load: n_mels        = 80
whisper_model_load: ftype         = 1
whisper_model_load: qntvr         = 0
whisper_model_load: type          = 2
whisper_model_load: mem required  =  310.00 MB (+    6.00 MB per decoder)
whisper_model_load: adding 1608 extra tokens
whisper_model_load: model ctx     =  140.66 MB
whisper_model_load: model size    =  140.54 MB
whisper_init_state: kv self size  =    5.25 MB
whisper_init_state: kv cross size =   17.58 MB

system_info: n_threads = 4 / 20 | AVX = 1 | AVX2 = 1 | AVX512 = 0 | FMA = 1 | NEON = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 1 | SSSE3 = 0 | VSX = 0 | COREML = 0 | OPENVINO = 0 |

main: processing 'C:\Users\qianp\Downloads\whisper.cpp_build-fix\bin\Release\Coldplay - Viva La Vida (Official Video).wav' (3873646 samples, 242.1 sec), 4 threads, 1 processors, lang = en, task = transcribe, timestamps = 1 ...


[00:00:00.000 --> 00:00:10.000]   [MUSIC]
[00:00:10.000 --> 00:00:16.000]   Ice to the world
[00:00:16.000 --> 00:00:19.000]   Seas it rise when I gave the world
[00:00:19.000 --> 00:00:23.000]   Now in the morning I sleep alone
[00:00:23.000 --> 00:00:27.000]   Sweet the streets I used to
[00:00:27.000 --> 00:00:37.000]   [MUSIC]
[00:00:37.000 --> 00:00:45.000]   I used to roll the dice
[00:00:45.000 --> 00:00:48.000]   Feel the fear in my enemies eyes
[00:00:48.000 --> 00:00:52.000]   Listen as the crowd would sing
[00:00:52.000 --> 00:00:55.000]   Now the old king is dead along with the king
[00:00:55.000 --> 00:00:59.000]   And in it I felt the key
[00:00:59.000 --> 00:01:02.000]   Mixed the walls with clothes don't mean
[00:01:02.000 --> 00:01:06.000]   And I discovered that my castle's done
[00:01:06.000 --> 00:01:10.000]   Upon pillars of sun, pillars of sun
[00:01:10.000 --> 00:01:13.000]   I can't ever stand the bells or ringin'
[00:01:13.000 --> 00:01:17.000]   Roman coverly cries of sin
[00:01:17.000 --> 00:01:21.000]   Be my mirror of my soul and see you
[00:01:21.000 --> 00:01:24.000]   Missionaries in the front of the field
[00:01:24.000 --> 00:01:27.000]   For some reason I can't explain
[00:01:27.000 --> 00:01:31.000]   Once you've gone, it was never, never
[00:01:31.000 --> 00:01:34.000]   All this work
[00:01:34.000 --> 00:01:37.000]   That was when I hold on
[00:01:37.000 --> 00:01:47.000]   [MUSIC]
[00:01:47.000 --> 00:01:54.000]   It was the wicked and wild win
[00:01:54.000 --> 00:01:57.000]   Blue down the doors to let me in
[00:01:57.000 --> 00:02:01.000]   The shattered windows and the sound of drums
[00:02:01.000 --> 00:02:05.000]   People couldn't believe what I'd become
[00:02:05.000 --> 00:02:09.000]   And revolutionaries wait
[00:02:09.000 --> 00:02:12.000]   For my head on a silver plate
[00:02:12.000 --> 00:02:16.000]   Just to prop it on a list of them
[00:02:16.000 --> 00:02:19.000]   Of who would ever wanna be killed
[00:02:19.000 --> 00:02:23.000]   My head's ever seen a bells or ringin'
[00:02:23.000 --> 00:02:27.000]   Roman coverly cries of sin
[00:02:27.000 --> 00:02:30.000]   Be my mirror of my soul and see you
[00:02:30.000 --> 00:02:34.000]   Missionaries in the front of the field
[00:02:34.000 --> 00:02:37.000]   For some reason I can't explain
[00:02:37.000 --> 00:02:40.000]   Once you've gone, it was never, never
[00:02:40.000 --> 00:02:43.000]   All this work
[00:02:43.000 --> 00:02:47.000]   That was when I hold on
[00:02:47.000 --> 00:03:15.000]   [MUSIC]
[00:03:15.000 --> 00:03:19.000]   It's ever seen a bells or ringin'
[00:03:19.000 --> 00:03:22.000]   Roman coverly cries of sin
[00:03:22.000 --> 00:03:26.000]   Be my mirror of my soul and see you
[00:03:26.000 --> 00:03:29.000]   Missionaries in the front of the field
[00:03:29.000 --> 00:03:32.000]   For some reason I can't explain
[00:03:32.000 --> 00:03:36.000]   Once you've gone, it was never, never
[00:03:36.000 --> 00:03:39.000]   All this work
[00:03:39.000 --> 00:03:46.000]   That was when I hold on
[00:03:46.000 --> 00:04:01.000]   [MUSIC]
[00:04:01.000 --> 00:04:11.000]   [BLANK_AUDIO]


whisper_print_timings:     load time =    81.76 ms
whisper_print_timings:     fallbacks =   0 p /   0 h
whisper_print_timings:      mel time =   146.55 ms
whisper_print_timings:   sample time =  1817.77 ms /  2250 runs (    0.81 ms per run)
whisper_print_timings:   encode time =  8818.34 ms /    12 runs (  734.86 ms per run)
whisper_print_timings:   decode time =  9967.03 ms /  2194 runs (    4.54 ms per run)
whisper_print_timings:    total time = 20885.13 ms

[bobqianic]

A very interesting phenomenon. Actually, that piece of code you had before, the one whose purpose you didn't know, has some practical significance. Its purpose is to enable Whisper to transcribe lyrics correctly, rather than outputting a bunch of [MUSIC].

Yes, but I believe the values are identical and this step is basically a noop. Needs confirmation. I guess I've put it there due to some doubts I had back when working on ggwave and was new to audio processing, but indeed the spectrum should be symmetrical.

https://github.com/ggerganov/whisper.cpp/blob/206fc93396936725bd362c93796cfdc8a87f8509/whisper.cpp#L2158-L2163
Last time when I modified log_mel_spectrogram, I removed it according to OpenAI's standards, which led to the lyrics not being transcribed. That's why the phenomenon described above occurred. We should add a fallback option, making ours the only Whisper implementation that can transcribe lyrics. @ggerganov

[ggerganov]

It's hard to make conclusions based on individual examples. It's easy to make the transcription output lyrics or not by playing with the parameters (like offset, prompt, context, etc.). Since the model is auto-regressive, small variations can lead to drastically different results.

I think the new version that does not sum the 2 parts of the spectrum is correct.
We did verify that this is what OpenAI Whisper also produces for the log mel spectrogram, correct?

[bobqianic]

We did verify that this is what OpenAI Whisper also produces for the log mel spectrogram, correct?

Yes, the new version is a perfect match.

[bobqianic]

It's hard to make conclusions based on individual examples.

So, let's run some more tests. Let's compare the output for a thousand random songs.

[bobqianic]

Since the model is auto-regressive, small variations can lead to drastically different results.

You are correct! I just came across this paper and used music as the prompt, which resolved the issue.

C:\Users\qianp\Downloads\whisper.cpp_build-fix\bin\Release>C:\Users\qianp\Downloads\whisper.cpp_build-fix\bin\Release\main.exe -m C:\Users\qianp\Downloads\whisper.cpp_build-fix\bin\Release\ggml-model-whisper-base.bin -f "C:\Users\qianp\Downloads\whisper.cpp_build-old\bin\Release\Coldplay - Viva La Vida (Official Video).wav" -t 4 -bs 5 -bo 5 --prompt music
whisper_init_from_file_no_state: loading model from 'C:\Users\qianp\Downloads\whisper.cpp_build-fix\bin\Release\ggml-model-whisper-base.bin'
whisper_model_load: loading model
whisper_model_load: n_vocab       = 51865
whisper_model_load: n_audio_ctx   = 1500
whisper_model_load: n_audio_state = 512
whisper_model_load: n_audio_head  = 8
whisper_model_load: n_audio_layer = 6
whisper_model_load: n_text_ctx    = 448
whisper_model_load: n_text_state  = 512
whisper_model_load: n_text_head   = 8
whisper_model_load: n_text_layer  = 6
whisper_model_load: n_mels        = 80
whisper_model_load: ftype         = 1
whisper_model_load: qntvr         = 0
whisper_model_load: type          = 2
whisper_model_load: mem required  =  310.00 MB (+    6.00 MB per decoder)
whisper_model_load: adding 1608 extra tokens
whisper_model_load: model ctx     =  140.66 MB
whisper_model_load: model size    =  140.54 MB
whisper_init_state: kv self size  =    5.25 MB
whisper_init_state: kv cross size =   17.58 MB

system_info: n_threads = 4 / 20 | AVX = 1 | AVX2 = 1 | AVX512 = 0 | FMA = 1 | NEON = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 1 | SSSE3 = 0 | VSX = 0 | COREML = 0 | OPENVINO = 0 |

main: processing 'C:\Users\qianp\Downloads\whisper.cpp_build-old\bin\Release\Coldplay - Viva La Vida (Official Video).wav' (3873646 samples, 242.1 sec), 4 threads, 1 processors, lang = en, task = transcribe, timestamps = 1 ...


[00:00:00.000 --> 00:00:13.000]   [music]
[00:00:13.000 --> 00:00:17.000]   Ice to the world
[00:00:17.000 --> 00:00:20.000]   Seas of rice when I gave the word
[00:00:20.000 --> 00:00:24.000]   Now in the morning I sleep alone
[00:00:24.000 --> 00:00:28.000]   Sweet the streets I used to
[00:00:28.000 --> 00:00:41.000]   [music]
[00:00:41.000 --> 00:00:45.000]   Ice to roll the dice
[00:00:45.000 --> 00:00:48.000]   Feel the fear in my enemies eyes
[00:00:48.000 --> 00:00:52.000]   Listen as the crowd would sing
[00:00:52.000 --> 00:00:56.000]   Now the old king is dead along with the king
[00:00:56.000 --> 00:00:59.000]   And in it I felt the key
[00:00:59.000 --> 00:01:02.000]   Mixed the walls with clothes don't mean
[00:01:02.000 --> 00:01:06.000]   And I just don't look like a castle stone
[00:01:06.000 --> 00:01:10.000]   Upon pillars of sun, pillars of sun
[00:01:10.000 --> 00:01:14.000]   I can't ever stand the bells or ringin'
[00:01:14.000 --> 00:01:17.000]   Roman coverly cries of sin
[00:01:17.000 --> 00:01:21.000]   Be my mirror, my soul, and see you
[00:01:21.000 --> 00:01:25.000]   Missionaries in the front of the field
[00:01:25.000 --> 00:01:28.000]   For some reason I can't explain
[00:01:28.000 --> 00:01:32.000]   Once you've gone, I was never, never old
[00:01:32.000 --> 00:01:38.000]   It's the word that was when I hold on
[00:01:38.000 --> 00:01:50.000]   [music]
[00:01:50.000 --> 00:01:54.000]   It was the wicked and wild wind
[00:01:54.000 --> 00:01:57.000]   Blue down the doors to let me in
[00:01:57.000 --> 00:02:01.000]   And the shattered windows and the sound of drums
[00:02:01.000 --> 00:02:05.000]   People couldn't believe what I'd be coming
[00:02:05.000 --> 00:02:09.000]   And revolutionaries wait
[00:02:09.000 --> 00:02:12.000]   For my head, I still have pain
[00:02:12.000 --> 00:02:16.000]   Just to prop it on a listing
[00:02:16.000 --> 00:02:19.000]   Of who would ever wanna be killed
[00:02:19.000 --> 00:02:23.000]   I can't ever stand the bells or ringin'
[00:02:23.000 --> 00:02:27.000]   Roman coverly cries of sin
[00:02:27.000 --> 00:02:30.000]   Be my mirror, my soul, and see you
[00:02:30.000 --> 00:02:34.000]   Missionaries in the front of the field
[00:02:34.000 --> 00:02:37.000]   For some reason I can't explain
[00:02:37.000 --> 00:02:41.000]   Once you've gone, I was never, never old
[00:02:41.000 --> 00:02:48.000]   It's the word that was when I hold on
[00:02:48.000 --> 00:03:15.000]   [music]
[00:03:15.000 --> 00:03:25.000]   Be my mirror, my soul, and see you
[00:03:25.000 --> 00:03:29.000]   Missionaries in the front of the field
[00:03:29.000 --> 00:03:32.000]   For some reason I can't explain
[00:03:32.000 --> 00:03:37.000]   Once you've gone, I was never, never old
[00:03:37.000 --> 00:03:43.000]   It's the word that was when I hold on
[00:03:43.000 --> 00:04:01.000]   [music]
[00:04:01.000 --> 00:04:11.000]   [BLANK_AUDIO]


whisper_print_timings:     load time =    83.76 ms
whisper_print_timings:     fallbacks =   0 p /   0 h
whisper_print_timings:      mel time =   145.07 ms
whisper_print_timings:   sample time =  1774.39 ms /  2253 runs (    0.79 ms per run)
whisper_print_timings:   encode time =  8038.97 ms /    11 runs (  730.82 ms per run)
whisper_print_timings:   decode time = 10365.85 ms /  2201 runs (    4.71 ms per run)
whisper_print_timings:    total time = 20461.74 ms

[NielsMayer]

FYI. Whisper itself suppresses "fun" tokens

openai/whisper#1839 (comment)