Deep Learning Practice · Quiz 2 · May 2026

Language identification depends on linguistic evidence in the speech signal. Background noise can mask phonetic cues; a very short clip may contain too little evidence; code-switching can mix languages within one labelled example. A large, balanced and representative training set normally supports better learning rather than being an adverse factor by itself. **Answer: A, B, C.** A: **Correct:** Can obscure the sound patterns needed for language recognition. B: **Correct:** May contain too few informative phonetic or lexical cues. C: **Correct:** Mixes language evidence and can violate a single-language assumption. D: **Incorrect:** Size and balance are generally beneficial when data quality and relevance are maintained.

Build speech regions with VAD/segmentation, transcribe them with timestamps, compute speaker embeddings, cluster those embeddings into speaker identities, and align the speaker intervals with the recognised text. Segment merging combines compatible neighbouring intervals into readable speaker-attributed turns. Language identification can be useful in multilingual settings, but it is not universally required to assign speakers and transcribe a known-language recording. TTS generates speech rather than analysing this recording. **Answer: A, B, C, E, G.** A: **Correct:** Defines speech regions and excludes irrelevant silence/non-speech. B: **Correct:** Supplies the transcript and its timing for speaker attribution. C: **Correct:** Provides representations from which speaker identity can be compared. D: **Incorrect:** Optional for this task, rather than a necessary diarization stage. E: **Correct:** Groups speech segments belonging to the same speaker. F: **Incorrect:** Speech synthesis is not needed to produce speaker-labelled text. G: **Correct:** Associates text/times with speaker segments and merges compatible turns.

The specified ECAPA-TDNN checkpoint is trained for mono 16 kHz speech. Its inference interface takes waveform batches and performs its configured acoustic-feature extraction internally. Temporal pooling turns a variable-duration speech sequence into a fixed-width speaker representation; duration can vary although extremely short clips may yield less reliable embeddings. Inputs still require suitable channel handling, sampling rate and valid lengths. **Answer: A, C, D.** A: **Correct:** Matches the checkpoint's single-channel training setup. B: **Incorrect:** Manual MFCC extraction is not required by the waveform inference interface. C: **Correct:** 16 kHz is the checkpoint's expected waveform sampling rate. D: **Correct:** Temporal pooling yields a fixed embedding dimension for variable-duration inputs.

Batch preparation must distinguish padding from signal preprocessing. Padded token targets use the loss's ignore index, conventionally -100. Dynamic padding uses the longest relevant sequence in the current batch instead of the whole dataset. SpeechT5 target lengths must be compatible with its reduction factor; implementations may round/truncate valid target lengths to a multiple and mask padded frames. Thus D describes the required multiple-of-r compatibility, not a universal promise that every implementation pads upward. Resampling and acoustic-feature extraction belong in the preprocessing pipeline and are not automatic responsibilities of a padding-only collator. **Answer: A, B, D.** A: **Correct:** Ignored padded token positions must not contribute to cross-entropy loss. B: **Correct:** Batch-local padding reduces unnecessary work for variable-length sequences; individual model constraints may still impose fixed feature sizes. C: **Incorrect:** A padding collator does not generally resample raw waveforms automatically. D: **Correct:** The target/frame layout must respect the decoder reduction factor; details can include rounding and truncation as well as padding. E: **Incorrect:** DataCollatorCTCWithPadding pads prepared inputs/labels; it is not a log-Mel feature extractor.

Sampling rate is the number of samples per second. Therefore \(N=f_sT=16000\times8=128000\) samples per channel. The question supplies one recording stream and does not ask for bytes; storage size would additionally depend on channel count and sample precision. **Answer: 128000.**

The requested tensor has one 1024-component vector for each of 64 clips, so its shape is (64,1024). Number of scalar entries \(=64\times1024=65536\). The 40 output classes and the full dataset's size do not multiply this pre-classifier mini-batch embedding tensor. **Answer: 65536.**

Word error rate is \(WER=(S+D+I)/N\), with N the number of reference words. The reference has 10 words. One minimum-edit alignment substitutes quick→fast and jumps→jumped (S=2), deletes the second 'the' before lazy (D=1), and inserts today (I=1). Thus WER=(2+1+1)/10=0.40. Count the reference words in the denominator, even if the hypothesis happens to have the same total length. **Answer: 0.40.**

last_hidden_state has a time dimension containing one feature vector per frame. mean(dim=1) averages over time while preserving batch and hidden-feature dimensions, giving one fixed-width vector per clip. In padded batches the mean should normally exclude padded frames using an appropriate mask. This question intends mean pooling as a duration-normalised clip representation. **Answer: C.** A: **Incorrect:** A sum also removes the time dimension, but its magnitude grows with the number of frames; it is not the intended normalised mean-pooling completion. B: **Incorrect:** PyTorch max(dim=1) ordinarily returns values and indices, not the plain mean-pooled embedding tensor required by this line. C: **Correct:** Averages frame vectors to produce the intended fixed-width clip embedding. D: **Incorrect:** Flattening combines time and feature dimensions, so the vector size varies with clip length.

compute_type selects the inference engine's numerical representation/quantisation mode. int8 uses 8-bit integer quantisation for supported model operations/weights, reducing memory compared with full precision. Accuracy and speed effects depend on hardware, model and data; a small accuracy change is typical motivation, not a mathematical guarantee. This setting does not change the audio file format, clip duration or number of recognised languages. **Answer: C.** A: **Incorrect:** The numeral refers to numerical precision, not an eight-second duration limit. B: **Incorrect:** The model's compute representation is separate from waveform file encoding. C: **Correct:** Describes the intended reduced-precision quantised inference mode. D: **Incorrect:** Language support is not truncated to eight languages by this setting.

vocab_size must be an integer matching the tokenizer's number of token IDs, including any added tokens; len(tokenizer) supplies that count. pad_token_id must be the numeric ID, not the padding token's string. These are the recognised configuration argument names for the CTC model. A supplies both with the correct types. **Answer: A.** A: **Correct:** Uses the correct argument names and numeric values. B: **Incorrect:** Uses unsupported names and provides token/string or dictionary values instead of the required IDs/count. C: **Incorrect:** The configuration names are pad_token_id and vocab_size, not pad_token and size. D: **Incorrect:** pad_token is a string and vocab_len is not the required configuration field.

CTC sums over possible frame-level alignment paths that collapse to the desired token string. The blank symbol lets a frame contribute no output label; the collapse operation removes blanks and merges consecutive repeated labels. This permits many frame sequences to map to a shorter text sequence without a provided frame-by-frame alignment. Blank is an alignment symbol, not necessarily a physically silent audio frame. **Answer: C.** A: **Incorrect:** Blank is not an end-of-sentence marker. B: **Incorrect:** It is not a replacement for unknown vocabulary words. C: **Correct:** It supports the alignment and collapsing of variable-length frame and label sequences. D: **Incorrect:** A blank prediction need not mean acoustic silence; it means no emitted label at that alignment step.

conv1 outputs 128 channels. Pooling changes sequence length but not channel count, so conv2 must accept 128 input channels. The subsequent shape calculation uses 256 output channels. Because the data are channel×sequence tensors, use Conv1d, with kernel 3, stride 1 and padding 1 to preserve sequence length through this convolution. **Answer: C.** A: **Incorrect:** Uses the original input_dim=1 instead of the 128 channels produced by conv1. B: **Incorrect:** Uses a 2D convolution and the wrong input-channel count. C: **Correct:** Conv1d(128,256,3,stride=1,padding=1) matches the incoming tensor and required output depth. D: **Incorrect:** Uses Conv2d although the architecture and tensor use one spatial/temporal dimension. E: **Incorrect:** Expects 256 incoming channels although conv1 supplies 128, and reverses the output depth.

Input shape is (100,1,768). conv1 uses kernel 3, stride 1, padding 1, so length stays 768 and channels become 128. The first size-2 stride-2 pool reduces length to 384. conv2 preserves length and produces 256 channels; the second pool reduces length to 192. Flattening each example therefore gives \(256\times192=49152\) features. The batch size 100 is preserved separately and must not multiply in_features. **Answer: 49152.**

A speaker-embedding waveform interface expects a 2D tensor shaped (batch,time). A stereo waveform loaded as (2,num_samples) has the same rank, so this interface interprets its first dimension as two batch examples. It produces one embedding for each channel rather than merging them automatically. For one mono representation, first apply the intended channel mixing/downmixing procedure and preserve an explicit batch dimension. **Answer: C.** A: **Incorrect:** The shape can be valid as a batch of two mono waveforms, so mono training alone does not imply a shape error here. B: **Incorrect:** The interface does not infer from this 2D shape that the rows should be downmixed. C: **Correct:** The first dimension is interpreted as batch, producing two embeddings. D: **Incorrect:** The embedding feature dimension is fixed; extra rows increase batch size instead.

A small fine-tuning set can give noisy gradients that rapidly overwrite useful pretrained representations. Freezing the encoder initially preserves those representations while the task-specific part adapts, reducing the risk of catastrophic forgetting and excessive fitting to the small dataset. Layers can be unfrozen later if validation supports it. **Answer: C.** A: **Incorrect:** Freezing training updates does not by itself remove encoder computation during inference. B: **Incorrect:** The training objective, such as CTC, remains necessary. C: **Correct:** Protects useful pretrained features from being overwritten too quickly. D: **Incorrect:** It does not add model parameters or capacity.

Characters are smaller units than typical BPE subwords. The same transcript therefore usually requires more output tokens with a character tokenizer, increasing sequence lengths and potentially alignment or decoder work. A character vocabulary is generally smaller and can spell new words from known characters; it is not inherently unable to represent unseen word combinations. **Answer: A.** A: **Correct:** The same text normally takes more character tokens than subword tokens. B: **Incorrect:** Character vocabularies are usually smaller, not larger. C: **Incorrect:** Known characters can compose words absent from the training vocabulary. D: **Incorrect:** A smaller vocabulary often reduces embedding-table size; this is not the primary disadvantage.

The acoustic model predicts an intermediate representation such as a Mel-spectrogram. The vocoder converts that time-frequency representation into waveform samples. Duration prediction determines how long units last; phonemisation converts text to pronunciation units; attention supports alignment. None of those replaces the waveform-generation role of the vocoder. **Answer: A.** A: **Correct:** Converts acoustic/spectral features to a time-domain signal. B: **Incorrect:** Predicts durations rather than synthesising the final waveform. C: **Incorrect:** Converts text into phonetic units. D: **Incorrect:** Aligns or weights representations rather than directly performing vocoding.

Agglomerative clustering repeatedly merges clusters according to a linkage criterion. Single linkage uses the smallest interpoint distance between two clusters. Ward linkage chooses a merge by the increase in within-cluster squared error and is normally paired with Euclidean geometry. Web linkage is not a standard criterion, and hierarchical describes the family of clustering methods rather than one particular linkage rule. **Answer: A, C.** A: **Correct:** Ward is a recognised agglomerative linkage criterion. B: **Incorrect:** Web is not a standard linkage criterion. C: **Correct:** Single linkage is a recognised criterion based on minimum intercluster distance. D: **Incorrect:** Hierarchical names the clustering framework, not a particular linkage definition.

There are \(400\times25=10000\) sentence vectors. Each contains 768 scalar values, so the array has \(10000\times768=7680000\) floating-point values. This asks for the number of values, not memory in bytes; byte size would additionally depend on whether storage uses float32, float64 or another dtype. **Answer: 7680000.**