Are there any local ASR models that surpass Whisper right now?
This covers generation capability or on-device inference progress — worth tracking for model efficiency, deployment cost, and application openings.
A user is seeking local Automatic Speech Recognition (ASR) models that outperform Whisper, specifically faster-whisper (medium/large turbo), while operating within an 8GB VRAM limi…
Hey everyone,
I'm currently using faster-whisper(medium/large turbo) for local speech recognition, running it on an 8GB VRAM GPU. It works great, but I was wondering if there are any new open-source/local models that outright beat Whisper at this point?
Here is exactly what I'm looking for in an alternative:
- Better Accuracy: Hoping for something with a noticeably lower WER (Word Error Rate) than Whisper.
- Multilingual Support: It needs to support at least English, Japanese, and Korean reliably.
- Speed: The inference speed should be comparable to, or ideally faster than, faster-whisper.
- Timestamps: It absolutely must support accurate timestamp outputs.
Also, since I'm constrained by 8GB of VRAM, it needs to be able to run within that limit. What is everyone using for local ASR these days? Any new SOTA recommendations? Thanks in advance!