2 wavs file with same content generate different result about faster-whisper

[Purfview]

0.wav and 1.wav are of same audio content

Not same, 1.wav actually has more audio content, you can see that from the file size.

Anyway, both files produced exactly same text:

  Processing segment at 00:00.000
[00:00.000 --> 00:01.800]  Let's talk about how you actually build a portfolio.
[00:02.020 --> 00:03.380]  I like to split it up in two ways.
[00:03.540 --> 00:06.360]  You have a core portfolio and you have a casino portfolio.
[00:06.520 --> 00:09.920]  Now the idea is that if you completely blow up your casino portfolio,
[00:10.100 --> 00:11.880]  you still have the majority intact.
[00:12.000 --> 00:14.020]  Over here, you can be a little bit more flexible.
[00:14.120 --> 00:18.120]  Now, I don't like gambling in crypto because gambling just inevitably leads to losses.
[00:18.360 --> 00:23.500]  You want to focus having the majority of your portfolio revolved around growing your Bitcoin and Ethereum holdings.
[00:23.560 --> 00:29.200]  The way that you increase how much Bitcoin that you have is by trading high conviction altcoins.
  Processing segment at 00:29.200
[00:29.200 --> 00:35.080]  But you should never let this whole core portfolio of Bitcoin and Ethereum shift 100% into altcoins.
[00:35.280 --> 00:40.100]  Because then you've gone all in at the poker table on a bet that you have no edge on winning.

generate different result

Model by default is non-deterministic. You would want to check DEBUG to see if it hits fallback, or just use temperature=0 in comparisons.

[leohuang2013]

What I mean, same content, is that, you play both audio file, what you hear is same.

I figure out what happened before. Different result was caused by different parameter passed to transcribe() function. with word_timestamps enabled/disabled, the result is different.

from faster_whisper import WhisperModel
model_size = "small"
model = WhisperModel(model_size, device="cpu", compute_type="int8")
audio = 'wrong_transcript.wav'
timestamp = True
segments, info = model.transcribe(audio, word_timestamps=timestamp, beam_size=5)
print("Detected language '%s' with probability %f" % (info.language, info.language_probability))
for segment in segments:
    print("[%.2fs -> %.2fs] %s" % (segment.start, segment.end, segment.text))

with word_timestamps enabled, that is, timestamp = True
output:

Detected language 'id' with probability 0.787822
[0.00s -> 3.44s]  Saya, Ibu Janu, Ibu Janu, selamat datang di Taut.
[3.44s -> 8.94s]  Saya tempat mengalirkan kesempatan RK10, RK10.8 dan perubatan.
[8.94s -> 10.44s]  Selamat datang di Taut.
[14.94s -> 18.28s]  Wah, RK10.8 ini dibetulah kamu sudah kena jadwanya.
[18.28s -> 20.84s]  Untuk banyak anak murid, orang pria harus melakukannya.
[20.84s -> 22.44s]  Semangat bawa-wawang.
[112.44s -> 114.44s]  Terima kasih.

with word_timestamps disabled, that is, timestamp = False
output:

Detected language 'id' with probability 0.787822
[0.00s -> 3.44s]  Saya, Ibu Janu, Ibu Janu, selamat datang di Taut.
[3.44s -> 8.94s]  Saya tempat mengalirkan kesempatan RK10, RK10.8 dan perubatan.
[8.94s -> 10.44s]  Selamat datang di Taut.
[14.94s -> 18.28s]  Wah, RK10.8 ini dibetulah kamu sudah kena jadwanya.
[18.28s -> 20.84s]  Untuk banyak anak murid, orang pria harus melakukannya.
[20.84s -> 22.44s]  Semangat bawa-wawang.
[52.44s -> 54.44s]  Kami di Kerala cuma lebih.
[59.44s -> 61.44s]  Kalau kandungnya di swanko,
[61.44s -> 64.44s]  yang menentang kandungnya adalah dengan baby A.
[64.44s -> 66.44s]  Kalau kandungnya banyak pribadi,
[66.44s -> 70.44s]  dan jadi lebih menakannya lebih jauh
[70.44s -> 73.44s]  dengan kodipi yang terkenal dari
[73.44s -> 74.44s]  sekalah SMA,
[74.44s -> 75.44s]  sekalah pun ya,
[75.44s -> 76.44s]  kandungnya lebih jauh.
[76.44s -> 79.44s]  Kami di Kerala cuma lebih.
[81.44s -> 83.44s]  Kalau kandungnya banyak pribadi,
[83.44s -> 85.44s]  dan kandungnya banyak pribadi,
[85.44s -> 87.44s]  dan kandungnya menjadi anak-anak
[87.44s -> 89.44s]  atau anak-anak už ke мне,
[89.44s -> 91.44s]  atau anak-anak terus main,
[91.44s -> 93.88s]  dan kemudian lagi nggak bisa
[94.44s -> 96.44s]  dari posisi dari kandungnya,
[96.44s -> 98.44s]  karena kandungnya seperti itu,
[98.44s -> 101.44s]  dan kemudian lagi,
[103.44s -> 105.44s]  Kandungnya ada sebulan.
[105.44s -> 110.44s]  Alas 10.000 yang dipakai sebanding namanya BANIL
[110.44s -> 112.44s]  Terima kasih dengan Pemerintah-Pemerintah

[leohuang2013]

I forgot attach wrong_transcript.wav file.
sample.zip

[Purfview]

What if you try medium model?