TL;DR My peer review workflow moves from understanding the paper to checking its claims, then writing and justifying a recommendation. The principle behind it is simple: ask what I would want to hear if I were the author. Useful feedback explains what I noticed, why it matters, and what needs clarification. The same habits help me review reports and plans at work. TL;DR 1. Why I’m writing this 2. One principle, three questions 3. The routine, step by step Step 0. Check the reviewing conditions Step 1. Read section by section, with notes Step 2. Check the frame Step 3. Identify what is different Step 4. Audit the numbers and settings Step 5. Cross-check the references Step 6. Flag strong language throughout Step 7. Summarize, then judge 4. From notes to a review What I try to make each comment do 5. The other side of the table: feedback I received 6. Where tools fit and what remains my responsibility 7. Beyond the conference 8. My closing checklist References 1. Why I’m writing this A paper represents months, sometimes years, of work. A reviewer has much less time to understand it and respond. I keep that imbalance in mind: a review may be brief, but it should never be careless. I received a Best Reviewer Award at Interspeech 2026, which recognized roughly the top 5% of reviewers. I review for Interspeech and ICASSP as an outside volunteer, on my own time. That recognition prompted me to reflect on my routine and share it, for two reasons: For peer reviewers : a practical starting point for reading a paper, checking its claims, and turning observations into useful feedback. For work : how the same habits can help us review reports and plans, with care for both the evidence and the people behind it. The Interspeech reviewer handbook [1] sets out the criteria. Here, I focus on applying them. If you review papers, technical reports, or proposals, I hope this is a useful starting point. I describe my process only, not the contents of any paper I reviewed. 2. One principle, three questions Before thinking about scores, I ask myself: If I were the author, what would I want to hear? I would want the reviewer to understand what I was trying to do, recognize what worked, and explain what did not. I would not want empty praise, but I would not want my effort dismissed either. That is the balance I aim for: respecting the authors does not mean agreeing with their claims, and being honest does not require being rude or harsh. That principle guides how I write. To assess the paper, I turn Interspeech’s three reviewing criteria [1] into three questions: Originality: What does the work add beyond what is already known? Technical correctness: Do the methods and evidence support the conclusions? Clarity: Can readers understand what was done and what was found? I keep these questions separate. A clear explanation does not make a result correct, and an unfamiliar idea is not necessarily new. 3. The routine, step by step My reviewing workflow. Watching for overclaims is an ongoing check, not a separate pass. Step 0. Check the reviewing conditions Before reading, I check the venue’s rules on confidentiality, anonymity, conflicts of interest, and tool use. I do not try to identify anonymous authors. If I have a conflict or cannot assess the work competently, I contact the meta-reviewer rather than guessing. Step 1. Read section by section, with notes I read and take notes without scoring yet, first understanding the problem, approach, evidence, and conclusion. My notes separate questions the paper answers later from concerns that remain. Step 2. Check the frame Is the problem clear? Is the objective coherent? Does the conclusion answer the question introduced at the beginning? Comparing the opening with the conclusion tells me where to look more closely. I also consider the type of contribution, since methods, datasets, replications, and surveys demonstrate their value differently. Step 3. Identify what is different I look at the main idea, figures, and prior work. What is different about the method, resource, analysis, or setting? “Interesting to me” is not the same as “original.” Complexity or a strong benchmark number does not prove novelty either, and I judge the project the authors presented, not the one I wish they had done. Step 4. Audit the numbers and settings I check whether the tables, comparisons, and experimental conditions support the claims. Are the baselines appropriate? Are important settings described? Do the reported results match the explanation in the text? The question is not only whether a number is impressive. It is what that number demonstrates under the conditions studied. Step 5. Cross-check the references I compare the work with existing papers and check whether the cited sources exist and support the claims they are cited for. I have encountered references I could not find. If I question originality, I also need to cite specific prior work rather than rely on a vague memory. Step 6. Flag strong language throughout Words such as “always,” “eliminates,” or “best” make me pause. They may be justified, but they require evidence matching their scope. My rule is simple: if you claim it, show the proof. I hold my own review to the same standard. Step 7. Summarize, then judge I summarize the paper in my own words to test whether I can explain its contribution without repeating the abstract. Only then do I settle on scores and a recommendation, taking a stand instead of defaulting to the middle of the scale, as the handbook asks [1]. The review should explain the reasoning, not just announce the outcome. 4. From notes to a review I use the handbook’s four-part structure [1] to organize the writing: Part of the review What I put here Summary My understanding of the problem, approach, and main findings Major strengths and weaknesses Contributions and concerns that affect the central claims Minor issues Presentation details, missing explanations, and inconsistencies Justification Why my assessment leads to the recommendation I look for genuine strengths rather than trying to fill a quota. When a concern matters, I explain it directly instead of hiding it behind polite but vague language. For presentation issues, I focus on helping readers understand the work, not making the paper sound as though I wrote it. What I try to make each comment do Make praise as specific as criticism. Rather than stopping at “well written” or “strong results,” I explain what helped me understand the work or what the analysis contributes. Tie each concern to something checkable. I identify a claim, comparison, or inconsistency, explain its consequence, and say what would clarify it: a clearer explanation, additional evidence, or a narrower claim. This guides the authors without making my recommendation conditional on a revision. Distinguish a limitation from a central flaw. A paper need not solve every related problem. Does the issue undermine the contribution or define its boundaries? That matters more than how many concerns I can list. The pattern I aim for is: What I observed, why it matters, and what would clarify it. For example (an invented one), instead of “The evaluation is insufficient, therefore the author's claim is too weak” I try to write: “The introduction claims robustness across recording conditions, but Table 2 reports one condition. That matters because the central claim is broader than the evidence. Could the authors narrow the claim or add a second condition?” For the overall recommendation, I add one question: how much does the issue change my confidence in the central contribution? 5. The other side of the table: feedback I received Some of the most useful reviews I received felt more like a discussion than a verdict, and two points in particular stayed with me. One concerned how I framed the problem: my wording could make a difficult problem sound simpler than it really was, while the setting I used to simulate it could give readers a misleading impression. The other questioned whether I had enough evidence to call a design choice the best option and suggested an ablation to test whether that claim was justified. What helped me move forward was not simply being told what was wrong, but understanding why it mattered and how I could address it. That experience reinforced two habits I now bring to my own reviews: looking closely at how a problem is framed and asking whether the evidence supports the claims. 6. Where tools fit and what remains my responsibility Tool use starts with the venue’s rules, not with what a tool can do. Interspeech 2026’s policy [1][2] says manuscripts must never be uploaded to public or commercial GenAI platforms, and that use is permitted only on firewall-protected or institutionally approved systems that guarantee confidentiality. The reviewer form further limits AI assistance to grammar editing and polishing text written by the reviewer. GenAI cannot assess originality, novelty, or correctness, its outputs are not scientific evidence, any use must be acknowledged in the review, and the reviewer stays fully responsible. I followed these rules. I used GenAI only to edit grammar and polish the review summary I had written myself, never to assess a paper or decide a recommendation. The system was firewall-protected and did not use inputs for training, and I acknowledged the use in my reviews. In the workplace context, our Research team has written about AI as a proofreader (in Japanese) [3]: useful for typos and logical slips, but limited by missing context and unreliable references. An internal workflow is not permission to apply the same tools to a confidential submission. Where allowed, tools may help with mechanics. Understanding the contribution, weighing the evidence, and explaining the judgment remain the reviewer’s responsibility. 7. Beyond the conference At work, I use many of the same habits on reports and plans: check the problem statement, ask whether the evidence supports the claims, identify real strengths, and keep the feedback respectful. For a plan, there are no results to audit yet. Instead, I ask whether the expected result is clear, what evidence would support it, and what outcome would show that the idea did not work. 8. My closing checklist Before submitting a review, I ask: Did I follow the confidentiality and anonymity rules? Can I explain the contribution in my own words? Did I separate originality, correctness, and clarity? Are my concerns tied to evidence, and is my praise specific? Did I distinguish central issues from limitations and minor details? Does my recommendation follow from the written assessment? Would I find this feedback fair and useful if I were the author? For me, a good review is not the one with the most criticism or the most generous praise. It helps the authors and decision-makers understand what the work establishes, where uncertainty remains, and why the recommendation follows. That is what I try to offer, and what I hope to receive when it is my turn to be reviewed. References Interspeech 2026 Reviewer Handbook, “Reviewing instructions.” https://wiki-is.isca-speech.org/en/Technical-Programme/Reviewer-handbook/Reviewing-instructions ISCA, “Code of Conduct for Reviewers and Meta-Reviewers.” https://isca-speech.org/https/www.isca-speech.org/Code-of-Conduct-for-Reviewers-and-Meta-Reviewers RevComm Tech Blog (in Japanese), “心理的安全性を高め、工数を減らす「校正者」としてのAIレビュー導入実践事例,” Dec 2025. https://tech.revcomm.co.jp/use-of-ai-in-review-process