Deep Learning Practice Quiz 2: 12 April 2026 (January 2026 term)

The tensor has shape: When using a pre-trained Transformer-based model (such as Wav2Vec 2.0) for a downstream classification task like Language Identification, what is the primary purpose of applying Mean Pooling across the sequence dimension?

1. A

   To reduce the sampling rate of the audio to make it compatible with standard deep learning layers.

2. B

   To convert a variable-length sequence of hidden vectors into a single fixed- length representation that summarizes the entire utterance.

3. C

   To ensure that the model focuses only on the most intense amplitude peaks of the speech signal.

4. D

   To reverse the effects of the Convolutional Neural Network (CNN) encoder and return the data to the time domain.

1. A

   Use before passing the audio through the model.

2. B

   Wrap the extraction code block with the : context manager.

3. C

   Manually set each layer's attribute to False using a for loop before every forward pass.

4. D

   Use the function on the output tensor to remove unnecessary dimensions.

1. A

   It is a standalone model that converts audio directly into text before passing it to the WhisperForConditionalGeneration class.

2. B

   It is a wrapper class that combines a WhisperFeatureExtractor (for audio) and a WhisperTokenizer (for text) into a single object to simplify data preprocessing and padding.

3. C

   It is primarily used to change the sampling_rate of the raw audio files to 16,000 Hz during the dataset.map() phase.

4. D

   It is a specialized loss function that calculates the Word Error Rate (WER) by comparing the model's audio features directly against the ground-truth text label

1. A

   To translate the text into different languages before speaking.

2. B

   To define the specific vocal characteristics (voice identity) of the speaker.

3. C

   To remove background noise from the generated file.

4. D

   To increase the speed of the audio generation.

Select all that apply.

1. A

   Speaker Embedding Extraction: Generating numerical representations (e.g., x- vectors) that capture the identity-bearing features of the vocal signal.

2. B

   Part-of-Speech (POS) Tagging: Assigning grammatical categories to words to determine the syntactic structure of the conversation.

3. C

   Clustering: Applying algorithms like Spectral Clustering or AHC to group embeddings into clusters corresponding to individual speakers.

4. D

   Named Entity Recognition (NER): Extracting proper nouns and specific entities from the transcribed text to identify participants by name.

5. E

   Distance/Similarity Metric Computation: Calculating the mathematical relationship (e.g., Cosine similarity or distances) between different speech segments.

Select all that apply.

1. A

   As the automatic speech recognition backend for transcription after diarization

2. B

   To provide word-level timestamps for aligning speaker segments

3. C

   As the primary clustering algorithm for speaker embeddings

4. D

   As the speaker embedding extractor

Select all that apply.

1. A

   Use .to(device) to move models to GPU

2. B

   Normalize audio waveforms before feeding to the embedding model

3. C

   Use batch processing for extracting embeddings from multiple segments

4. D

   Pass raw text tokens to the speaker encoder

5. E

   Use torch.no_grad() during inference to save memory

You are building an ASR pipeline for Assamese. Complete the missing parts of the script to ensure the audio is resampled, features are extracted, and the trainer is initialized correctly.

Based on the above data, answer the given subquestions.

Choose the option from below to fill in the blank at A.

1. A

   apply

2. B

   cast_column

3. C

   transform

4. D

   batch_decode

You are building an ASR pipeline for Assamese. Complete the missing parts of the script to ensure the audio is resampled, features are extracted, and the trainer is initialized correctly.

Based on the above data, answer the given subquestions.

Choose the option from below to fill in the blank at B.

1. A

   WhisperTokenizer

2. B

   WhisperProcessor

3. C

   WhisperForConditionalGeneration

4. D

   Seq2SeqTrainingArguments

You are building an ASR pipeline for Assamese. Complete the missing parts of the script to ensure the audio is resampled, features are extracted, and the trainer is initialized correctly.

Based on the above data, answer the given subquestions.

Choose the option from below to fill in the blank at C.

1. A

   feature_extractor

2. B

   tokenizer

3. C

   collator

4. D

   None of these

output of print(vocabs) is as given below

Based on the above data, answer the given subquestions.

For the given code what needs to be filled in place of missing code to get the vocabulary which consists of only unique character ?

1. A

   

2. B

   —

3. C

   

4. D

   

1. A

   To translate the input text into a different language.

2. B

   To tokenize the text into sub-word units.

3. C

   To convert intermediate acoustic features (like mel-spectrograms) into audible waveforms.

4. D

   To compress the final audio file into a ZIP format.

1. A

   To make the file size as large as possible.

2. B

   Because Python cannot process any frequency higher than 16kHz.

3. C

   Because 16kHz is the highest quality audio humans can hear.

4. D

   Because the model was specifically trained on 16kHz audio data.