Acoustic Speech Recognition with Hyper-Networks

(1) Baseline Model (Single Utterance)

• Stacked LSTM with single utterance train setting
• Test model with single utterance
• Test model with two concatenated utterances

(2) Baseline Model (Two Utterances)

• Stacked LSTM with two utterances train setting
• Test model with single utterance
• Test model with two concatenated utterances

(3) Baseline Model (Two Utterances) using I-Vector

• Considering I-Vector as addtional condition data
• Stacked LSTM with two utterances train setting
• Test model with single utterance
• Test model with two concatenated utterances

(4) Hyper Network Model (Two Utterances) without Regularizers

• Using hyper-networks
• Stacked LSTM with two utterances train setting
• Test model with single utterance
• Test model with two concatenated utterances

(5) Hyper Network Model (Two Utterances) with Regularizers (Slow-Feature)

• Using hyper-networks with additional cost terms (slow feature)
• Stacked LSTM with two utterances train setting
• Test model with single utterance
• Test model with two concatenated utterances

Goal is to learn user-adaptive features, and this feature decides the model by generating parameters.