The IIT Madras BS Deep Learning Practice (Deep Learning Practice) Quiz 2 paper sat on 3 Aug 2025, in the May 2025 term: 11 questions for 50 marks in 120 minutes. Every question is below with its answer. Take it as a timed mock test to be marked, or read it through first.
In a speaker diarization pipeline that uses Whisper for transcription and an embedding model for speaker identification,you observe that one speaker's segments are consistently broken into multiple smaller segments, each assigned a different speaker label (e.g., SPEAKER 1, SPEAKER 3, SPEAKER 5). The transcribed text for this speaker is perfectly accurate. What is the MOST likely cause of this specific issue?
Correct answer
Consider the following code snippet for using a pretrained Wav2Vec2 model:
from transformers import Wav2Vec2Processor, Wav2Vec2ForCTCimport torch
processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-base-960h")model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-base-960h")
# Pretend this is audio datainput_values = torch.randn(16000) # 1 second of fake audio
# Forward passlogits = model(input_values).logitsWhat is the primary issue in the above code?
Correct answer
Which statements accurately describe the roles and differences of the data collators used in the ASR and TTS training scripts?
Correct answers
Correct answers
You are calculating the Word Error Rate (WER) for an ASR model's prediction. Given the following ground truth and prediction, what is the resulting WER? Provide the answer rounded to two decimal places.
from jiwer import wer
ground_truth = "i like to eat apples and bananas"prediction = "i like eating apple and banana"
error = wer(ground_truth, prediction)print(error)Correct answer: 0.57 (accepted within ±0.02)
Correct answer: 1536000
Consider the following code snippet that processes a raw audio sample first with a Wav2Vec2Processor and then with a Wav2Vec2Model.
from transformers import Wav2Vec2Processor, Wav2Vec2Modelimport torch
processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-base-960h")model = Wav2Vec2Model.from_pretrained("facebook/wav2vec2-base-960h")# A 1-second audio sample at 16kHzsample_audio = torch.randn(1, 16000)
# Method 1: Processing the audioinputs_processed = processor(sample_audio, sampling_rate=16000, return_tensors="pt")# The shape of inputs_processed['input_values'] is (1, 16000)
# Method 2: Passing processed inputs to the modelwith torch.no_grad(): outputs_model = model(inputs_processed['input_values']) features = outputs_model.last_hidden_state# The shape of features is (1, 49, 768)Based on the above data, answer the given subquestions.
To convert the raw audio waveform into a sequence of contextualized hidden states.
To normalize the raw audio waveform and prepare it in the correct tensor format for the model.
To downsample the audio from its native sampling rate to 16000 Hz.
To tokenize the audio into a sequence of discrete units, similar to a text tokenizer.
Correct answer
To normalize the raw audio waveform and prepare it in the correct tensor format for the model.
Consider the following code snippet that processes a raw audio sample first with a Wav2Vec2Processor and then with a Wav2Vec2Model.
from transformers import Wav2Vec2Processor, Wav2Vec2Modelimport torch
processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-base-960h")model = Wav2Vec2Model.from_pretrained("facebook/wav2vec2-base-960h")# A 1-second audio sample at 16kHzsample_audio = torch.randn(1, 16000)
# Method 1: Processing the audioinputs_processed = processor(sample_audio, sampling_rate=16000, return_tensors="pt")# The shape of inputs_processed['input_values'] is (1, 16000)
# Method 2: Passing processed inputs to the modelwith torch.no_grad(): outputs_model = model(inputs_processed['input_values']) features = outputs_model.last_hidden_state# The shape of features is (1, 49, 768)Based on the above data, answer the given subquestions.
The model's output features has a sequence length of 49, while the input audio had a length of 16000 samples. Based on these values, what is the approximate downsampling factor applied by the Wav2Vec2Model's feature encoder (the CNN part)? Provide the answer rounded to the nearest integer.
Correct answer: 327
Consider the following code snippet that processes a raw audio sample first with a Wav2Vec2Processor and then with a Wav2Vec2Model.
from transformers import Wav2Vec2Processor, Wav2Vec2Modelimport torch
processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-base-960h")model = Wav2Vec2Model.from_pretrained("facebook/wav2vec2-base-960h")# A 1-second audio sample at 16kHzsample_audio = torch.randn(1, 16000)
# Method 1: Processing the audioinputs_processed = processor(sample_audio, sampling_rate=16000, return_tensors="pt")# The shape of inputs_processed['input_values'] is (1, 16000)
# Method 2: Passing processed inputs to the modelwith torch.no_grad(): outputs_model = model(inputs_processed['input_values']) features = outputs_model.last_hidden_state# The shape of features is (1, 49, 768)Based on the above data, answer the given subquestions.
Correct answers
You are fine-tuning a Wav2Vec2ForCTC model for an ASR task using the Hugging Face Trainer. The following TrainingArguments are used:
from transformers import TrainingArguments
training_args = TrainingArguments( output_dir="AsrTaskModel", group_by_length=True, per_device_train_batch_size=8, evaluation_strategy="steps", num_train_epochs=7, fp16=True, save_steps=1000, eval_steps=500, learning_rate=1e-4, warmup_steps=1000, save_total_limit=2, load_best_model_at_end=True,)Based on the above data, answer the given subquestions.
To sort the entire dataset by audio duration before training begins.
To ensure each batch contains audios of exactly the same length, eliminating the need for padding.
To create batches with samples of similar input length, which minimizes the amount of padding required and improves training efficiency.
To group samples by the length of their text transcription rather than their audio length.
Correct answer
To create batches with samples of similar input length, which minimizes the amount of padding required and improves training efficiency.
You are fine-tuning a Wav2Vec2ForCTC model for an ASR task using the Hugging Face Trainer. The following TrainingArguments are used:
from transformers import TrainingArguments
training_args = TrainingArguments( output_dir="AsrTaskModel", group_by_length=True, per_device_train_batch_size=8, evaluation_strategy="steps", num_train_epochs=7, fp16=True, save_steps=1000, eval_steps=500, learning_rate=1e-4, warmup_steps=1000, save_total_limit=2, load_best_model_at_end=True,)Based on the above data, answer the given subquestions.
Correct answers