3 ms·

Isn't phonetics already a solved problem in ASR? The hard part is NLU, including semantics and pragmatics, which this article mainly talks about.

StudentStuff 8y ago
Some languages are significantly easier to train against than English, for example 80hrs of Turkish got some decent results for this group using Mozilla Deepspeech: https://arxiv.org/pdf/1807.00868.pdf https://arxiv.org/pdf/1807.00868.pdf For context, Baidu used a bit over 5000 hours of English to get a decent model. Also, OP may have a pronunciation difference (regional, hyper-local or just from learning English as a second language) that isn't handled well by Google's models.
- conistonwater 8y ago
  I suspect deep learning researchers have no real incentive to use the minimum amount of data sufficient to get their thing working. The worst case for them is if their model underperforms because of insufficient data instead of something truly DL-related, so they use as much data as they have/can handle. It could still be that you're right, but I wouldn't rely on those kinds of published numbers.
timonoko 8y ago
Whatever, but it does not work, as I can easily demonstrate. I once tried to buy flashlight batteries from Vancouver. I tried every possible variation of BATTERY, but had to write it on paper. Problem is that English does not have standard pronunciation, because writing is not phonetic. In other languages everybody knows what is the official pronunciation even if their own dialect vary wildly.
- Kiro 8y ago
  Just tried saying battery in many different ways, even mumbling and it got it correctly every time. No idea how you pronounce stuff in order to have such problems.
  - timonoko 8y ago
    The problem at time was the stress point. In most of my languages the stressing is flat or in the back. At first I started using the standard "BAT-TE-RII" of almost all languages, but realized soon it is maybe pronounced "BAT-RII", but failed to emphasize the first syllabel as in "BATTT-ri" and ended up varying forms of "baaat-RRRIIIIII".

StudentStuff 8y ago

Some languages are significantly easier to train against than English, for example 80hrs of Turkish got some decent results for this group using Mozilla Deepspeech: https://arxiv.org/pdf/1807.00868.pdf https://arxiv.org/pdf/1807.00868.pdf For context, Baidu used a bit over 5000 hours of English to get a decent model. Also, OP may have a pronunciation difference (regional, hyper-local or just from learning English as a second language) that isn't handled well by Google's models.

conistonwater 8y ago
I suspect deep learning researchers have no real incentive to use the minimum amount of data sufficient to get their thing working. The worst case for them is if their model underperforms because of insufficient data instead of something truly DL-related, so they use as much data as they have/can handle. It could still be that you're right, but I wouldn't rely on those kinds of published numbers.

timonoko 8y ago

Whatever, but it does not work, as I can easily demonstrate. I once tried to buy flashlight batteries from Vancouver. I tried every possible variation of BATTERY, but had to write it on paper. Problem is that English does not have standard pronunciation, because writing is not phonetic. In other languages everybody knows what is the official pronunciation even if their own dialect vary wildly.

Kiro 8y ago
Just tried saying battery in many different ways, even mumbling and it got it correctly every time. No idea how you pronounce stuff in order to have such problems.
- timonoko 8y ago
  The problem at time was the stress point. In most of my languages the stressing is flat or in the back. At first I started using the standard "BAT-TE-RII" of almost all languages, but realized soon it is maybe pronounced "BAT-RII", but failed to emphasize the first syllabel as in "BATTT-ri" and ended up varying forms of "baaat-RRRIIIIII".