Logo Lanfrica

Beniam-Kumela/amharic-asr-ipynb

Domaine:

natural language processing

Type de record:

modelsoftware
Créateur:
Ben
Hôte:
# amharic-asr-ipynb This repo implements a DeepSpeech2-like speech recognition model and trains it on an Amharic dataset. This is meant to host the code discussed in this video. We stress that the main purpose of this model is as a teaching resource as the implementation is more beginner-friendly than the current state-of-the-art (SOTA). Specifically, we find that the final model has a character error rate (CER) of ~23% which is worse than the SOTA trained on the same dataset which achieved a CER of ~7%. The model only takes ~20 epochs (which totals to 2-3 hrs of training time on Google Colab GPUs) to converge as shown by the loss curves (data included in repo). Some predictions from the model are shown below. ``` True : እንደ ገና ኢትዮጵያዊ ለ መሆን መደራደር የ መረጡ ም አሉ Pred: እንደ ገና ኢትዮጵያዊ ለ መሆን መደራድር የ መረጡ ህሉ CER: 11.11111111111111 % True : ስለ ጦርነቱ መናገር በ እነሱ ስለሚ ያምር ጉዳዩ ን ለ እነሱ መተው በ ፈቀድኩ Pred: ስለ ጥርምነቱ መናገር በ እነሱ ስለሚየመር ጉዳዩ ን ለነሱ መተው በፈቀም CER: 20.408163265306122 % True : ለ ዜጐቻቸው የተደላደለ ኑሮ ፈጥረው ከ ተለያዩ ሀገሮች የ መጡ ስደተኞች ን ያስተናግ ዳሉ በ ማለት በልዩ አይ ን እንመለከታ ቸው ይሆናል Pred: ለ ዜጐቻቸው የተደላደረ ኑሮ ፈጥረው ከ ተለያዩ አገሮች የ መቶ ስደተኞች ን ያስተናግዳሉ በ ማለት መልዩዮ አይን ን መልከታቸው ይሆናል CER: 12.790697674418606 % ``` Please feel free to open any pull requests if there are any problems with running this code or if there are additional functionalities you would like to see implemented.