I'm working on the same project myself and was planning to write a blog post similar to the author's. However, I'll share some additional tips and tricks that really made a difference for me.
For preprocessing, I found it best to convert files to a 16kHz WAV format for optimal processing. I also add low-pass and high-pass filters to remove non-speech sounds. To avoid hallucinations, I run Silero VAD on the entire audio file to find timestamps where there's a speaker. A side note on this: Silero requires careful tuning to prevent audio segments from being chopped up and clipped. I also use a post-processing step to merge adjacent VAD chunks, which helps ensure cohesive Whisper recordings.
For the Whisper task, I run Whisper in small audio chunks that correspond to the VAD timestamps. Otherwise, it will hallucinate during silences and regurgitate the passed-in prompt. If you're on a Mac, use the whisper-mlx models from Hugging Face to speed up transcription. I ran a performance benchmark, and it made a 22x difference to use a model designed for the Apple Neural Engine.
For post-processing, I've found that running the generated SRT files through ChatGPT to identify and remove hallucination chunks has a better yield.
If I understood correctly, VAD has superior results than using ffmpeg silencedetect + silentremove, right?
I think latest version of ffmpeg could use whisper with VAD[1], but I still need to explore how with a simple PoC script
I'd love to know more about the post-processing prompt, my guess is that looks like an improved version of `semantic correction` prompt[2], but I may be wrong ¯\_(ツ)_/¯ .
Since the past two days I've been working on SpeechShift [1], its a fully local, offline first, speech to text utility that allows you to trigger it with a command, transcribes with whisper and puts pastes it in the window you are currently focused on (like chrome, typora or some other window). Basically SuperWhisper [2] but for linux. (If this is something which interests you & check it out! Feel free to ping me if something does not work as expected.)
I've been trying to squeeze out performance out of whisper, but felt (at least for non native speakers) the base model does a good job. In terms of pre processing I do VAD & some normalization. But on my rusty thinkpad the processing time is way too long. I'll try some of the forementioned tips and see if the accuracy & perf can get any better. Post which I'm planning to use a SLM for text cleanup & post processing of the transcription. I'm documenting my learnings over at my notes [3].
I was going to go the opposite way and suggest that if you want python audio transcription, you can skip ffmpeg and just use whisper directly. Using the whisper module directly gives you a variety of outputs, including text and srt.
Yep. Whisper is great. I use it on podcasts as part of removing ads. Last time I used one of the official versions it would only accept .wav files so I had to convert with ffmpeg first.
Nice job. I made a similar python script available as a Github gist [1] a while back that given an audio file does the following:
- Converts to 16kHz WAV
- Transcribes using native ggerganov whisper
- Calls out to a local LLM to clean the text
- Prints out the final cleaned up transcription
I found that accuracy/success increased significantly when I added the LLM post-processor even with modestly sized 12-14b models.
I've been using it with great success to convert very old dictated memos from over a decade ago despite a lot of background noise (wind, traffic, etc).
I forget all the details of my tweaks, but I remember that I had better throughput on my version.
I know the OP talked about wanting it local, but thomasmol/whisper-diarization on replicate is fast and cheap. Here's a hacked front end to parse teh JSON: https://github.com/Sanborn-Young/MP3_2transcript
I also have an app that does this fully locally and offline on the macOS app store; Wisprnote - using the openai whispr models. Works good.
What people are talking about, avoiding hallucinations through VAD based chunking, etc, are all things I pioneered with Wisprnote, which has been on the App Store for 2 years. Hasn't been updated recently - backlog of other work - but still works just as fine. Paid app. But good quality.
I personally love senko since it can run in seconds, whereas py-annote took hours, but there is a 10% WER (word error rate) that is tough to get around.
btw, if you want local dictation, speak and get a transcript, not transcribe files, I built a Python tool called hns [1]. It's open source, uses faster-whisper, and you can run it with `uvx hns` or just `hns` after `uv tool install hns`.
I was using the same setup to try to transcribe a sound track of a video. A 60s aac audio took me maybe 10 minutes. I'm on a apple M4 and ran `whisper audio.aac --model medium --fp16 False --language Japanese`. Wonder if I'm doing something wrong
For preprocessing, I found it best to convert files to a 16kHz WAV format for optimal processing. I also add low-pass and high-pass filters to remove non-speech sounds. To avoid hallucinations, I run Silero VAD on the entire audio file to find timestamps where there's a speaker. A side note on this: Silero requires careful tuning to prevent audio segments from being chopped up and clipped. I also use a post-processing step to merge adjacent VAD chunks, which helps ensure cohesive Whisper recordings.
For the Whisper task, I run Whisper in small audio chunks that correspond to the VAD timestamps. Otherwise, it will hallucinate during silences and regurgitate the passed-in prompt. If you're on a Mac, use the whisper-mlx models from Hugging Face to speed up transcription. I ran a performance benchmark, and it made a 22x difference to use a model designed for the Apple Neural Engine.
For post-processing, I've found that running the generated SRT files through ChatGPT to identify and remove hallucination chunks has a better yield.