Hacker News Reader

Comment by HackedBunny | Hacker News Reader

My quick test of DeepSpeech, as posted on Facebook back in April:

I just ran a speech-to-text converter on a very clear clip of former Doctor Who actor Tom Baker talking in an interview.

The DeepSpeech converter uses the very latest AI deep-learning advancements to 'listen' to the audio and output the spoken words as text.

After 3 long minutes of running it on a 30-second clip, it printed out its interpretation:

"hooloomooloo how booboorowie i have a honeymoon"

View on HN

It's supposed to work on sentence long audio (4 - 5 seconds), they suggest chunking your audio first: https://discourse.mozilla.org/t/longer-audio-files-with-deep...

Also there's essentially two parts to this, the neural net is used for speech-to-characters, and then a language model is used to convert the character stream to words.

I found that the language model they supplied was trained data that did not contain the words I needed, and got significantly improved results when making my own language model using the kenlm[1] tools.

[1]: https://kheafield.com/code/kenlm/

Would it be possible to substitute this for GPT2/BERT? Or is that a different type of language model? Can the pre-trained language model be fine-tuned? I’m using DeepSpeech to transcribe long-form lecture audio, and have just assumed there would be a massive improvement once they noise-harden the models with 1.0.

GPT2 is not a good language model but there are things like XLM. Mozilla DeepSpeech doesn't support XLM rescoring, other toolkits do and it gives great improvement in accuracy. If you care about accurate transcription you'd better consider alternatives.

I didn't know any other ML-based open source transcription engines existed? I can't seem to find them on Google.

I didn’t know Tom Baker was Welsh.

When I tried it out with English and German phone recordings, it was working competitively. I would have ranked it better than Amazon but worse than Google.

Did you maybe not convert your WAV to the correct sampling rate?

Experiment is worthless for drawing conclusions if it's not reproducible by other people.

Besides the hypothesis that DS sucks, the software could also very well be just fine and you made methodological errors.

Reply on news.ycombinator.com