Scientific Paper Annotated Bibliography

Template Overview

Conducting literature reviews across dozens of academic journals quickly leads to cognitive overload. An annotated bibliography maintains structured modular records of each paper, capturing central theoretical arguments, methodology strengths, limitations, and relevance to your research project in plain Markdown.

Markdown Template Structure

Copy the block below or click Download .md File to add this template directly to your personal vault:

---
title: "Annotated Bib: [Research Field / Project Topic]"
date_updated: [YYYY-MM-DD]
researcher: "[[Researcher / Student Name]]"
project_title: "Hardware-Accelerated Local Audio Processing"
total_sources: 12
primary_fields: ["Acoustics", "Machine Learning", "Information Privacy"]
tags: [annotated-bibliography, academic, literature-review, citations, research]
---

# Scientific Paper Annotated Bibliography — [Research Field]

## 1. Project Research Scope & Inclusion Criteria
- **Central Research Question:** How do quantized on-device transformer models compare in word error rate (WER) and thermal efficiency against server-grade architectures?
- **Inclusion Criteria:** Peer-reviewed conference proceedings (NeurIPS, ICASSP, ACL) and IEEE/ACM transactions published between 2023 and 2026.

## 2. Annotated Source Entries

### Entry 01: Foundational Acoustic Modeling
- **Citation (APA 7th):**
  > Radford, A., Kim, J. W., Xu, T., Brockman, G., McLeavey, C., & Sutskever, I. (2023). Robust Speech Recognition via Large-Scale Weak Supervision. *Proceedings of the 40th International Conference on Machine Learning (ICML)*, 28492–28518.
- **Keywords:** `speech recognition`, `weak supervision`, `zero-shot transfer`, `whisper`
- **Theoretical Contribution:** Demonstrates that scaling weakly supervised web audio datasets (680,000 hours) produces acoustic models that are exceptionally robust to background noise and varied accents without task-specific fine-tuning.
- **Methodology & Dataset:** Encoder-decoder transformer architecture trained on 30-second audio chunks with Mel-spectrogram input representations.
- **Methodological Limitations:** Model exhibits susceptibility to repetitive phrase loops during non-verbal pauses or prolonged audio silence.
- **Relevance to My Project:** Serves as the primary acoustic architecture for our local transcription pipeline; establishes baseline WER targets.

---

### Entry 02: Quantization & Edge Efficiency
- **Citation (APA 7th):**
  > Gerganov, G., et al. (2024). Porting Whisper to Native C/C++: Memory-Mapped Weights and SIMD Vectorization on ARM Architecture. *Journal of Open Source Software*, 9(98), 6420.
- **Keywords:** `quantization`, `ggml`, `apple-silicon`, `neon-simd`, `edge-ai`
- **Theoretical Contribution:** Proves that integer 4-bit and 8-bit quantization (Q4_K, Q8_0) achieves 3.8x faster inference on consumer ARM laptops with less than 0.8% degradation in transcription accuracy.
- **Methodological Limitations:** Evaluated primarily on desktop macOS; lacks comprehensive thermal throttling metrics under continuous multi-hour loads.
- **Relevance to My Project:** Provides direct mathematical and empirical justification for deploying local on-device transcription without cloud GPUs.

## 3. Comparative Literature Synthesis Matrix
| Source | Core Model | Quantization | Dataset Size | Reported WER | Local Feasibility |
|---|---|---|---|---|---|
| Radford et al. (2023) | Transformer Large | FP32 | 680k hrs | 4.2% | High RAM (Server) |
| Gerganov et al. (2024) | ggml whisper.cpp | Q4_K | Identical | 4.9% | **Excellent (Local Mac)** |

## 4. Identified Literature Gaps
> [!NOTE] Scholarly Gap
> Existing papers benchmark isolated short audio snippets (under 30 seconds), but lack comprehensive literature evaluating multi-hour technical meeting transcripts on battery power.

## 5. Next Sources to Acquire
- [ ] Procure ICASSP 2026 proceedings on streaming encoder latency.
- [ ] Review Apple Machine Learning Research papers on unified memory bandwidth.

Filled Real-World Example

Here is a realistic example demonstrating how this template captures actual deliberations:

---
title: "Annotated Bib: Differential Privacy in NLP"
date_updated: 2026-10-06
researcher: "[[Elena Rostova]]"
project_title: "Privacy-Preserving Local Note Generation"
total_sources: 8
tags: [annotated-bib, privacy, nlp]
---

# Scientific Paper Annotated Bibliography — Differential Privacy in NLP

## 1. Research Overview
Elena Rostova compiled an annotated bibliography examining $(\epsilon, \delta)$-differential privacy bounds applied to local text summarization models.

## 2. Key Synthesis
The literature confirms that running model inference entirely on local client hardware provides absolute information privacy $(\epsilon = 0)$, rendering cloud differential privacy noise injection unnecessary.

When to Use and When to Avoid

When to Use

• Use when initiating a thesis, grant proposal, or journal literature review.
• Use to maintain a persistent, searchable library of read academic papers.
• Use to verify citation accuracy and methodological rigor before publication.

When to Avoid

• Avoid using for fleeting casual book recommendations or fiction novels.
• Avoid writing superficial single-sentence summaries that skip methodology.

On-Device AI Integration

Summarize your recording in Oyma with this local instruction:

Analyze the PDF text or abstract of the academic paper:
1. Generate a formal APA 7th edition citation and extract 4 core keywords.
2. Summarize the central theoretical contribution, methodology, and empirical dataset.
3. Identify methodological limitations and format into an annotated bibliography Markdown entry.

Your data remains securely on your Mac with zero cloud exposure.