Spaces:

openai

/

whisper

like 1.81k

Running on L4

Hey @ImagineThis ! Welcome to the HuggingFace community! There's an updated space here that can transcribe audio samples of arbitrary length: https://huggingface.co/spaces/sanchit-gandhi/whisper-large-v2

It's not real-time, but you don't have limits on your audio input length :) We're looking into ways of making this faster!

Hope that helps!

fyi, I do use Whisper Cloud APIs for "kindof" realtime transcriptions. My short experience with it mirrors what Whisper docs mention:

1. Latency depends on model. I HAVE to use large for multilingual support. It is 3x-8x slower than small or tiny e.g
2. Latency depends on total #emitted tokens. So, you have to chunk your audio input.

Is it English you prefer using?

@sanjaymk908 Yes, English. I've found a few projects that have attempted real-time, but it's all done in Terminal and require a good Nvidia GPU, which I don't have (All Mac).

@sanchit-gandhi Thanks for the links. Hopefully, we will get a proper UI soon for real-time. It will put Dragon out of business.

Thanks both.

For English & Mac, this is a fantastic solution: https://github.com/ggerganov/whisper.cpp

I have validated its central claims on small & medium whisper models:

1. The stripped down models it creates+uses can be run on a mobile device (not useful for your usecase).
2. Its latencies are staggeringly low & optimized for Macs
3. All this is delivered in 2 cpp files (can be easily customized)

I haven't yet tested large models. Ensure you see this in stdout to confirm you indeed are running optimized on your MacBook:
system_info: n_threads = 4 / 8 | AVX = 0 | AVX2 = 0 | AVX512 = 0 | FMA = 0 | NEON = 1 | ARM_FMA = 1 | F16C = 0 | FP16_VA = 1 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 0 | VSX = 0 |

You can maybe also try "streaming" mode to start transcribing as soon as the first token is ready: https://twitter.com/joao_gante/status/1641435222679130112

@sanchit-gandhi Thanks for this. I presume this is using Whisper, but it's still not really useful for day to day use on Mac OS .

Ideally, I want to press a button and, with a Whisper app loaded in the background, I can use speech to text in any application.

There are other efforts, it's getting closer anyday. I would even pay someone to make what I just described, and I bet millions of people would love it.

Hey @ImagineThis - you can try using this API which is more than 50x faster than OpenAI's model: https://huggingface.co/spaces/sanchit-gandhi/whisper-jax

You might hit a queue though if there are people in front of you wanting to use the demo.

Have you tried this for MacOS: https://support.apple.com/en-gb/guide/mac-help/mh40584/mac

It states:

On a Mac with Apple silicon, dictation requests are processed on your device in many languages — no internet connection is required.

why not splitting models by languages like VOSK API?
it will increase loading speed significantly

I have created a quick sample live transcription script if you're interested: https://github.com/gaborvecsei/whisper-live-transcription

Super basic idea but based on a few subjective tests, it looked all right, even with the tiny model:

1st second: [t, 0, 0, 0] --> "Hi"
2nd second: [t-1, t, 0, 0] --> "Hi I am"
3rd second: [t-2, t-1, t, 0] --> "Hi I am the one"
4th second: [t-3, t-2, t-1, t] --> "Hi I am the one and only Gabor"
5th second: [t, 0, 0, 0] --> "How" --> Here we started the process again, and the output is in a new line
6th second: [t-1, t, 0, 0] --> "How are"
etc....

Here's an updated guide for running the model using your microphone in real-time: https://huggingface.co/learn/audio-course/chapter7/voice-assistant#speech-transcription

I had the same problem and so I’ve created a working proof of concept for real time transcription with a websocket server and a demo JS client here if you want to check it out
https://github.com/alesaccoia/VoiceStreamAI