Datasets:
The dataset viewer is not available for this split.
Error code: RowsPostProcessingError
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
EECS 183/283A HW4
Public student materials for the speech-processing homework. Release: hw4-2026-v4. Download a notebook from this repository and open it in Google Colab. Its setup cell downloads the support bundle automatically; later cells fetch only the relevant data and TTS starter. Public downloads need no Hugging Face account/token. The original HuBERT, CosyVoice, and Whisper models are fetched from their upstream Hugging Face repositories at the pinned revisions specified in the support bundle.
- Part A: probing HuBERT
- Part B: ASR
- Part C: TTS
- Student instructions
- Report template
- Data descriptions and sources
| Course file | Download size |
|---|---|
hw4-student.zip |
0.6 MiB |
hw4-data-a.zip |
226.4 MiB |
hw4-data-b.zip |
141.4 MiB |
hw4-data-c.zip |
118.1 MiB |
hw4-speakers.zip |
214.6 MiB |
hw4-tts-starter.pt |
406.7 MiB |
A/B use a compact LibriSpeech subset. C retains all 354,780 original tokenized training samples. The TTS starter contains the original 768-width/12-layer/12-head model's weights after two epochs. The notebook defaults to one further epoch. The small support ZIP excludes the large data/model files listed separately above.
For a manual download, use hf_hub_download(repo_id='Louis0324/eecs183-hw4', repo_type="dataset", revision='hw4-2026-v4', filename="hw4-data-a.zip", token=False).
The notebooks verify data/checkpoint checksums and reuse cached downloads.
release_manifest.json lists file sizes and SHA-256 hashes. Published release
tags are not moved; a revised course release gets a new version.
Submit one combined ZIP with one report.pdf to Gradescope, as explained in the notebooks. Do not submit datasets or model checkpoints. Staff solutions, private grading references, and the Gradescope autograder are not public assets.
Sources: A/B derive from the pinned LibriSpeech/MFA alignment source documented in data/README.md; attribution is retained inside their archives. C and speaker data preserve the previous homework's caches. Vendored code keeps its upstream source and license notices in the student bundle's vendor directory.
- Downloads last month
- 59