Q: Please clarify "unmatched accuracy" but then in one of your responses you noted...
You mentioned in another comment, you recommend to proofread the output. We don't speak the languages we're converting transcriptions to. Hence buying your product. So what do we do, if we buy and we have no idea how to proofread, not seeing the value that we then should go and check again somewhere else? Please kindly explain this for potential buyer like us...
Terry_SubEasy.ai
Nov 21, 2024A: Hi there! That’s an excellent question!
The primary reason for recommending proofreading is that you're creating a professional video, where precision matters. Proofreading ensures the final output is flawless.
However, if you're simply working on a hobby project or in situations where perfection isn’t crucial, you can leave the results as they are.
Our service provides the best automatic transcription and translation available. Particularly in translations, once the transcription is accurate, you can rely on our AI-powered translations to deliver excellent results. AI Plus translation processes the entire transcript comprehensively, ensuring consistency and understanding throughout.
Conversely, if you're transcribing a language you’re not familiar with, translating the transcript can help you gain a better understanding.
I hope this answers your question!
yumeconsultsPLUS
Verified purchaser
How about you give us a variance instead of using words like "unmatched accuracy" then turn around and say hobby project.. It's not clear. Most users here on AppSumo clearly not using this for hobby.., can you not do the marketing jargon and just give us a % of accuracy like other one's on here 99.9, or 99.8 like maybe that can help buyers... Please.
Sure. You can get results with like 0-3 words error for every minute depend on the audio language. Especial less error on roman languages and Chinese, Korean and Japanese.
I do not recommend taking on professional projects, such as agency work, in a language with which you are completely unfamiliar.
The hobby project I'm talking about is something like personal YouTube channel.
The point of our product is to get have unmatched accuracy(compared to other similar project) to reduce endless proofreading, but can't completely eliminate human work.
user3500PLUSVerified purchaser
You asked for number, and since I did several hour of testing to compare, here we go:
Audio: German, Expert speakers, recorded in a professional studio with broadcast microphone
Best other tool I know: Subtitle Edit with FasterWhisper XXL, default parameters, model large-v3turbo. Result: 5 wrong words for each 10 min of audio (about 2.000 words). No deletions, no insertions/hallucinations
user3500PLUSVerified purchaser
With subeasy using glossary for names, I get the best result:
9 errors for each 10 min/2000 words, BUT their "AI refine" catches about 7 of them (shows about 11, 2 are artistic choices, 2 are correct, 7 are errors. Takes me 60 seconds to check with refine).
2 errors are left in 2000 words/10 min, no deletions/insertions.
WER = 0.1%, accuracy 99.9% - for the specific HQ audio mentioned above
user3500PLUSVerified purchaser
@yumeconsults: WER (word error rate) , % + hallucinations numbers will ALWAYS vary — I tested about 30 tools during the last 2 years and spend weeks comparing Whisper models+options. It depends on the spoken language, the speaker, background noise, the exact words (medical/technical/informal etc). You have to test it for YOUR audio. SubEasy has a cool AI finetune which fixes most remaining errors.
user3500PLUSVerified purchaser
Studies claim lower WER + fewer hallucinations for English/German clean Audio with CrisperWhisper over large-v3-turbo. For my audio, it's the opposite. Studies claim FasterWisper would be less accurate - with large-v3turbo, but I still get the fastest + overall best results with it. Some speakers mispronunce words and a modell fixing it to make sense is what you want - but NOT for comedians...
Interesting to see you have less errors with Large-v3turbo models, we're still using the Large-V2 model for it's versatilities among most languages. We will do more tests when new stuff coming out.