Question 1
Which of the following factors can adversely impact the accuracy of speech language identification?
Background noise
Very short utterances
Code-switching
Large balanced training data
Deep Learning Practice Quiz 2: 16 August 2026 (May 2026 term)
The IIT Madras BS Deep Learning Practice (Deep Learning Practice) Quiz 2 paper sat on 16 Aug 2026, in the May 2026 term: 19 questions for 48 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.
- 19
- 48
- 120 min
- 5
- 5
- 9
Show answer
Correct answers
Background noise
Very short utterances
Code-switching
Question 2
You are building an end-to-end speaker-attributed transcription pipeline using the Whisper model for speech recognition and timestamp generation, the ECAPA-TDNN model for speaker embedding extraction, and the SpeechBrain toolkit to integrate the diarization pipeline.
Which of the following subtasks are necessary to achieve this? (Select all that apply)
Voice Activity Detection (VAD) / Segmentation
Speech Recognition and Timestamp Generation
Speaker Embedding Extraction
Language Identification
Speaker Clustering
Text-to-Speech (TTS) Synthesis
Alignment and Segment Merging
Show answer
Correct answers
Voice Activity Detection (VAD) / Segmentation
Speech Recognition and Timestamp Generation
Speaker Embedding Extraction
Speaker Clustering
Alignment and Segment Merging
Question 3
When feeding audio into the pretrained speechbrain/spkrec-ecapa-voxceleb model, which of the following statements regarding the expected input is (are) true? (Select all that apply)
The system is trained on single-channel (mono) recordings.
The user must manually extract acoustic features (like MFCCs or Mel- filterbanks) before passing the data to the model.
The expected sampling rate for the input audio is 16kHz.
The model accepts variable-length audio inputs, dynamically generating fixed- size embeddings regardless of the utterance duration.
Show answer
Correct answers
The system is trained on single-channel (mono) recordings.
The expected sampling rate for the input audio is 16kHz.
The model accepts variable-length audio inputs, dynamically generating fixed- size embeddings regardless of the utterance duration.
Question 4
Which statements accurately describe the roles, constraints, and data-processing behaviors when preparing batches for speech models (e.g., Whisper, Wav2Vec2, and SpeechT5) using the Hugging Face Transformers library?
Dynamic padding in speech data collators pads input features and target labels only to the maximum sequence length within each batch, reducing unnecessary memory usage compared to dataset-wide padding.
Speech data collators automatically resample raw audio (e.g., from 44.1 kHz to 16 kHz) during batch collation before creating model inputs.
Show answer
Correct answers
Dynamic padding in speech data collators pads input features and target labels only to the maximum sequence length within each batch, reducing unnecessary memory usage compared to dataset-wide padding.
Question 5
An audio recording is sampled at 16,000 Hz for 8 seconds. How many samples are contained in the recording?
Show answer
Correct answer: 128000
Question 6
A multilingual language identification model is trained to classify 40 languages. The training dataset contains 1000 audio clips per language, and each audio clip has been resampled to 16 kHz. The Transformer encoder produces a 1024 dimensional embedding for every audio clip.
During training, a mini-batch contains 64 audio clips. Before the classification layer, the embeddings are stacked into a tensor of shape (batch_size, embedding_dimension).
Question: Determine the total number of embedding values in this output tensor.
Show answer
Correct answer: 65536
Question 7
An Automatic Speech Recognition (ASR) system produces a hypothesis transcript for a spoken audio clip. Calculate the Word Error Rate (WER) of the system given the reference (ground truth) and hypothesis transcripts below: Reference (Ground Truth): "the quick brown fox jumps over the lazy black dog"• Hypothesis (ASR Output): "the fast brown fox jumped over lazy black dog today"• (Note: Provide your answer as a decimal rounded to two decimal places.)
Show answer
Correct answer: 0.4 (accepted within ±0.01)
Question 8
In a spoken language identification task, a pretrained speech model is used to extract features from an input audio waveform. To obtain a single fixed-length embedding representing the entire audio clip for language classification, fill in the missing statement in the following code.
sum
max
mean
flatten
Show answer
Correct answer
mean
Question 9
What is the effect of setting compute_type="int8" in WhisperModel(whisper_model, compute_type="int8")?
It restricts the model to only transcribing 8-second audio chunks.
It forces the audio file to be read as an 8-bit WAV file.
It quantizes the model weights to 8-bit integers, significantly reducing GPU/CPU memory usage with minimal accuracy loss.
It limits the language detection capability to the 8 most common languages.
Show answer
Correct answer
It quantizes the model weights to 8-bit integers, significantly reducing GPU/CPU memory usage with minimal accuracy loss.
Question 10
Consider the following incomplete script intended for fine-tuning a Wav2Vec2 model:
Which of the following code blocks correctly fills in the blank to ensure the model's output layer matches the tokenizer's vocabulary size and recognizes its padding token?
pad_token_id=tokenizer.pad_token_id, vocab_size=len(tokenizer)
padding_id=tokenizer.pad_token, vocabulary=tokenizer.vocab
pad_token=tokenizer.pad_token_id, size=tokenizer.vocab_size
pad_token_id=tokenizer.pad_token, vocab_len=len(tokenizer)
Show answer
Correct answer
pad_token_id=tokenizer.pad_token_id, vocab_size=len(tokenizer)
Question 11
Which of the following describes the correct input for the Decoder during the training of a Seq2Seq ASR model like Whisper?
The raw audio waveform normalized to a zero mean.
A sequence of random noise to improve robustness.
The Log-Mel Spectrogram extracted from the audio signal.
The tokens predicted by the decoder up to the previous time step during inference
Show answer
Correct answer
The tokens predicted by the decoder up to the previous time step during inference
Question 12
When fine-tuning a pre-trained Transformer-based ASR model using Connectionist Temporal Classification (CTC), what is the primary function of the 'blank' token?
Indicating the end of a sentence.
Handling out-of-vocabulary words.
Aligning variable-length sequences.
Representing silent audio segments.
Show answer
Correct answer
Aligning variable-length sequences.
Question 13
Based on the above data, answer the given subquestions.
The input tensor train_data has shape
Determine the correct value of in_features for the fc1 layer.
Show answer
Correct answer: 49152
Question 14
It will automatically mean-pool the left and right channels into a mono signal before processing.
It will treat the tensor's first dimension as a batch dimension, producing two separate embeddings — one per channel.
It will produce a single embedding, but with double the feature dimension to account for the extra channel.
Show answer
Correct answer
It will treat the tensor's first dimension as a batch dimension, producing two separate embeddings — one per channel.
Question 15
Why might you freeze the encoder layers of a pre-trained ASR model during the initial phases of fine-tuning on a very small dataset ?
Reducing the inference latency.
Eliminating the need for CTC.
Preventing catastrophic forgetting.
Increasing the model capacity.
Show answer
Correct answer
Preventing catastrophic forgetting.
Question 16
What is the primary disadvantage of using a character-level tokenizer compared to a Byte-Pair Encoding (BPE) subword tokenizer for ASR fine-tuning ?
Longer output sequence lengths.
Larger vocabulary size.
Inability to predict new words.
Higher memory usage for embedding.
Show answer
Correct answer
Longer output sequence lengths.
Question 17
In a standard TTS pipeline involving an acoustic model and a vocoder, which component is primarily responsible for converting Mel-spectrograms into time-domain waveforms during the fine-tuning process?
The Vocoder.
The Duration Predictor.
The Phonemizer.
The Attention Mechanism.
Show answer
Correct answer
The Vocoder.
Question 18
Which of the following linkages is possible in agglomerative clustering?
Ward Linkage
Web Linkage
Single Linkage
Hierarchical Linkage
Show answer
Correct answers
Ward Linkage
Single Linkage
Question 19
The "sentence-transformers/all-mpnet-base-v2" model generates 768-dimensional sentence embeddings. A data processing pipeline processes 400 unique documents, each containing 25 sentences. It stores the embedding vector for every sentence in a single contiguous NumPy array. How many floating-point values does the array contain?
Show answer
Correct answer: 7680000