uis-rnn gives different result on broken audios and continuous audios

[ashu170292]

Describe the question

I trained the uis-rnn model on embeddings obtained on timit data. I am calculating embedding over a 240 ms window with 50 % overlap. I am using this uis-rnn model to obtain speaker ids from real time audios. For this I am using TOEFL test audios which have 3-4 speakers per recording. Each recording is 50-60 seconds long

When I use the model on continuous audios, I get only one or two speaker ids. But, if I break down the audio corresponding to different speakers and concatenate the embeddings corresponding to the broken audios in a sequence, I get different and fairly accurate cluster ids. This would mean that the model is performing differently on continuous and broken audios. For the continuous audios, I tried different level of overlap and different number of embeddings per second , but no improvement.

Attached is the TOEFL audio and the same audio broken into three parts that I am using to test the model. The three parts broken audios are corresponding to different speakers in the continuous audioi
toefl_continuous_recording.wav.zip
broken_toefl_3_spk_recording.zip

My background

Have I read the README.md file?

• yes

Have I searched for similar questions from closed issues?

• yes

Have I tried to find the answers in the paper Fully Supervised Speaker Diarization?

• yes

Have I tried to find the answers in the reference Speaker Diarization with LSTM?

• yes

Have I tried to find the answers in the reference Generalized End-to-End Loss for Speaker Verification?

• yes

[wq2012]

How did you train your uis-rnn network?

Did you also train it on continuous audio?

[ashu170292]

Q)How did you train your uis-rnn network?
Answer) To train the uis-rnn network, I made a train sequence as a single 2 -dim numpy array. I used around ~4000 utterances of timit. Each utterance has only one speaker. Each utterance is 4 seconds to 6 seconds long. For each utterance, embeddings were calculated and was appended in the train sequence.

Q)Did you also train it on continuous audio?
Answer) No

[wq2012]

The whole point of UIS-RNN is to learn conversational information from examples. If your UIS-RNN is trained on single-speaker utterance only, the trained model will be useless on multi-speaker audio.

[ashu170292]

Thanks for your help, Quan.

The uis-rnn model trained on single speaker utterance, performs bad on multi-speaker utterance(This can be explained by the answer above). I am still finding it hard to build intuition around the following:

If I break down the multi speaker audio corresponding to different speakers and concatenate the embeddings corresponding to the broken audios in sequence, I get different and fairly accurate predicted ids (around 91 % accuracy).

Any idea why would that happen?

[wq2012]

UIS-RNN is an algorithm for supervised learning. This means, you train on multi-speaker data, it will perform well on multi-speaker data. You train it on single-speaker data only, it will only perform well on single-speaker data. It's not supposed to perform well on scenarios that never appeared during training.

When I use the model on continuous audios, I get only one or two speaker ids. But, if I break down the audio corresponding to different speakers and concatenate the embeddings corresponding to the broken audios in a sequence, I get different and fairly accurate cluster ids.

This seems unrelated to UIS-RNN. Sounds like a bug in your speaker embedding implementation. If you extract speaker embeddings from sliding windows, whether it is continuous audio or broken audio should not make much difference.