Hugging Face's logo

openai
/
whisper-large-v3

Automatic Speech Recognition
Transformers
PyTorch
JAX
Safetensors
whisper
audio
hf-asr-leaderboard
Eval Results
Model card Files
xet
 Community
242

100% Offline CPU-optimized Whisper STT implementation in a free Windows workstation

#242
by MDRstudio - opened
Discussion
MDRstudio

Hi community,

For anyone looking for a completely private, offline Whisper speech-to-text implementation on Windows without dealing with Python environments, cloud APIs, or token costs:

We integrated a local, CPU-optimized Whisper engine directly into KIRA Company Director (a free Windows 10/11 desktop workstation).

It generates accurate subtitles (.SRT, .VTT) and full transcripts completely locally on your machine with zero internet connection required.

Hope this helps anyone needing an out-of-the-box local Whisper desktop setup!
Organictruth

Do you have a reproducible offline CPU recipe for transcribing proper names and numeric identifiers with faster-whisper? Our tiny.en and base.en tests each made three word edits across four synthetic utterances. Vocabulary hints fixed a name but introduced unrelated words elsewhere. We preserve raw transcripts and timestamps. Which tested tools or decoding settings improve this without guessing corrections? A pinned example and test log would help. Thank you.
MDRstudio

Hi @Organictruth ,

To be completely transparent with you: in My experience building KIRA, completely eliminating hallucinations on smaller models like tiny and base is virtually impossible due to the nature of Whisper's sequence-to-sequence decoder.

However, the single biggest improvement we found isn't in post-processing or heavy prompt engineering, but in strict upstream audio preprocessing:

  1. Pre-processing to clean 16kHz Mono:
    Before feeding audio to Whisper, we strip all extra channels and resample via FFmpeg strictly to 16kHz single-channel PCM 16-bit (pcm_s16le, -ac 1, -ar 16000). Feeding unnormalized multi-channel audio or low-quality streams triggers most of the digit and name hallucinations.

  2. Noise Floor & Silence:
    Whisper tends to hallucinate vocabulary hints during low-volume or noisy pauses. Pre-filtering background noise and aggressive silence gating (or setting no_speech_threshold=0.6) helped us cut down unrelated token generation significantly.

  3. Condition on Previous Text:
    If you pass vocabulary hints, make sure condition_on_previous_text=False. Otherwise, if it slightly mishears a proper name once, that hallucination cascades across the entire rest of the transcript.

Beyond that, for strict numeric IDs and rare proper nouns, small.en is really the sweet spot where the model stops guessing.

Sign up or log in to comment