How to get timestamps per word instead of token

[kai-shimada]

Hello,

currently, there is a --max-length flag to get timestamps per token if set to 1 but it doesn't always output timestamps per word.

Is it possible to get timestamps per word?

Thanks!

This is my output:

./main -m ./models/ggml-small.bin -t 7 -f ./audio/223692__filmbrute__sony-pcm-m10-voice-test_s16le.wav -ml 1

whisper_model_load: loading model from './models/ggml-small.bin'
whisper_model_load: n_vocab       = 51865
whisper_model_load: n_audio_ctx   = 1500
whisper_model_load: n_audio_state = 768
whisper_model_load: n_audio_head  = 12
whisper_model_load: n_audio_layer = 12
whisper_model_load: n_text_ctx    = 448
whisper_model_load: n_text_state  = 768
whisper_model_load: n_text_head   = 12
whisper_model_load: n_text_layer  = 12
whisper_model_load: n_mels        = 80
whisper_model_load: f16           = 1
whisper_model_load: type          = 3
whisper_model_load: adding 1608 extra tokens
whisper_model_load: mem_required  = 1044.00 MB
whisper_model_load: ggml ctx size =  464.56 MB
whisper_model_load: memory size   =   68.48 MB
whisper_model_load: model size    =  464.44 MB

system_info: n_threads = 7 / 10 | AVX = 0 | AVX2 = 0 | AVX512 = 0 | NEON = 1 | ARM_FMA = 1 | F16C = 0 | FP16_VA = 1 | WASM_SIMD = 0 | BLAS = 1 | 

main: processing './audio/223692__filmbrute__sony-pcm-m10-voice-test_s16le.wav' (669042 samples, 41.8 sec), 7 threads, 1 processors, lang = en, task = transcribe, timestamps = 1 ...


[00:00:00.000 --> 00:00:00.000]  
[00:00:00.000 --> 00:00:00.450]   Hi
[00:00:00.450 --> 00:00:00.900]  ,
[00:00:00.900 --> 00:00:01.800]   this
[00:00:01.800 --> 00:00:02.250]   is
[00:00:02.250 --> 00:00:03.600]   George
[00:00:03.600 --> 00:00:04.270]   Der
[00:00:04.270 --> 00:00:04.720]  ug
[00:00:04.720 --> 00:00:05.320]  in
[00:00:05.320 --> 00:00:06.070]   from
[00:00:06.070 --> 00:00:06.970]   Film
[00:00:06.970 --> 00:00:07.420]  br
[00:00:07.420 --> 00:00:08.160]  ute
[00:00:08.160 --> 00:00:08.760]  .
[00:00:08.760 --> 00:00:09.530]  com
[00:00:09.530 --> 00:00:09.700]   and
[00:00:09.700 --> 00:00:10.000]   this
[00:00:10.000 --> 00:00:10.270]   is
[00:00:10.270 --> 00:00:10.400]   a
[00:00:10.400 --> 00:00:10.800]   mic
[00:00:10.800 --> 00:00:11.340]   test
[00:00:11.340 --> 00:00:11.760]   for
[00:00:11.760 --> 00:00:12.170]   the
[00:00:12.170 --> 00:00:12.720]   Sony
[00:00:12.720 --> 00:00:13.110]   PC
[00:00:13.110 --> 00:00:13.120]  M
[00:00:13.120 --> 00:00:13.280]   M
[00:00:13.280 --> 00:00:14.280]  10
[00:00:14.280 --> 00:00:16.450]   recorder
[00:00:16.450 --> 00:00:17.080]  .
[00:00:17.080 --> 00:00:17.150]   I
[00:00:17.150 --> 00:00:17.300]  'm
[00:00:17.300 --> 00:00:17.610]   just
[00:00:17.610 --> 00:00:17.990]   going
[00:00:17.990 --> 00:00:18.140]   to
[00:00:18.140 --> 00:00:18.370]   bab
[00:00:18.370 --> 00:00:18.600]  ble
[00:00:18.600 --> 00:00:18.980]   about
[00:00:18.980 --> 00:00:19.440]   random
[00:00:19.440 --> 00:00:19.840]   things
[00:00:19.840 --> 00:00:20.510]   and
[00:00:20.510 --> 00:00:20.820]   hopefully
[00:00:20.820 --> 00:00:21.050]   you
[00:00:21.050 --> 00:00:21.280]  'll
[00:00:21.280 --> 00:00:21.740]   listen
[00:00:21.740 --> 00:00:21.930]   to
[00:00:21.930 --> 00:00:22.250]   the
[00:00:22.250 --> 00:00:22.760]   quality
[00:00:22.760 --> 00:00:22.990]   of
[00:00:22.990 --> 00:00:23.450]   this
[00:00:23.450 --> 00:00:24.080]   audio
[00:00:24.080 --> 00:00:24.380]   and
[00:00:24.380 --> 00:00:24.610]   it
[00:00:24.610 --> 00:00:25.500]   will
[00:00:25.500 --> 00:00:25.530]   help
[00:00:25.530 --> 00:00:25.880]   you
[00:00:25.880 --> 00:00:27.150]   understand
[00:00:27.150 --> 00:00:27.310]   why
[00:00:27.310 --> 00:00:27.700]   this
[00:00:27.700 --> 00:00:28.250]   recorder
[00:00:28.250 --> 00:00:28.400]   is
[00:00:28.400 --> 00:00:29.000]   regarded
[00:00:29.000 --> 00:00:29.170]   so
[00:00:29.170 --> 00:00:29.490]   well
[00:00:29.490 --> 00:00:29.760]  .
[00:00:29.760 --> 00:00:30.490]  
[00:00:30.490 --> 00:00:31.280]   Right
[00:00:31.280 --> 00:00:32.180]   now
[00:00:32.180 --> 00:00:32.490]   I
[00:00:32.490 --> 00:00:33.110]  'm
[00:00:33.110 --> 00:00:33.690]   in
[00:00:33.690 --> 00:00:33.990]   a
[00:00:33.990 --> 00:00:35.510]   quiet
[00:00:35.510 --> 00:00:36.760]   room
[00:00:36.760 --> 00:00:37.280]   and
[00:00:37.280 --> 00:00:37.280]   the
[00:00:37.280 --> 00:00:37.710]   level
[00:00:37.710 --> 00:00:37.880]   is
[00:00:37.880 --> 00:00:38.140]   set
[00:00:38.140 --> 00:00:38.310]   to
[00:00:38.310 --> 00:00:38.570]   7
[00:00:38.570 --> 00:00:38.840]  .
[00:00:38.840 --> 00:00:38.840]  
[00:00:38.840 --> 00:00:39.600]   [
[00:00:39.600 --> 00:00:41.810]  BLANK
[00:00:41.810 --> 00:00:41.810]  _
[00:00:41.810 --> 00:00:41.810]  AUDIO
[00:00:41.810 --> 00:00:48.840]  ]


whisper_print_timings:     load time =   219.04 ms
whisper_print_timings:      mel time =    57.81 ms
whisper_print_timings:   sample time =    13.96 ms
whisper_print_timings:   encode time =  2236.72 ms / 186.39 ms per layer
whisper_print_timings:   decode time =   957.59 ms / 79.80 ms per layer
whisper_print_timings:    total time =  3518.50 ms