Exploring End-to-End Techniques for Low-Resource Speech Recognition — mentioned by StudentStuff

Mon, 10 Dec 2018 09:57:06 GMT

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 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.

Paper and discussion history. A mention is not an endorsement.

Exploring End-to-End Techniques for Low-Resource Speech Recognition · HackerPapers discussions

Exploring End-to-End Techniques for Low-Resource Speech Recognition — mentioned by StudentStuff