Analyze User Interview Recordings With AI: Build a Source-Checked Evidence Matrix

Define the Evidence Deliverable Before Asking for Themes

To analyze user interview recordings with AI, aim for an evidence matrix and a separate list of open research questions—not just a polished summary of what users supposedly want. Begin with one narrow question and at least two consented interview recordings saved in FoneClaw.

Our Voice Notes plugin provides saved recordings, transcripts, and collection context for questions across selected sources. The matrix described here is a proposed research method and a working artifact you review. It is not a dedicated native coding platform or an automated validation of your findings.

For example, ask how participants move information from a conversation into later action, rather than asking AI to determine a complete product strategy. That narrower question makes it easier to identify evidence that supports, limits, or contradicts an interpretation.

Define which interviews are included, why they are relevant, and who may receive the output. Scope any finding to those selected participants and situations. A collection answer does not establish statistical representation, causality, or market demand.

The useful endpoint is a claim you can trace back to speech, with uncertainty still visible—not a confident sentence detached from its source.

Select Permitted Recordings and Assign Neutral Labels

Confirm that recording, processing, the intended research purpose, and the proposed sharing are permitted. Include only the content needed for that purpose. Android AI Meeting Recording Consent: From Capture to Confirmed Actions provides a fuller consent and sharing workflow.

Inside a compatible Agent, Voice Notes supports recording, pausing, resuming, saving, and naming. Check each saved audio recording and its transcript before adding it to the analysis. Collections select at least two saved in-app recordings and reuse existing transcripts.

Assign neutral labels such as I01, I02, and I03, along with a session date and topic. Keep any necessary identity mapping separately with appropriate access. These are your labels; generated speaker labels do not automatically verify real-person identity.

For the full source-selection procedure, use FoneClaw Recording Collections: Build a Source-Checked Project Review. For individual recording preparation, AI Audio Summarizer for Android: Record, Transcribe, and Review Notes covers transcript and source review.

The plugin is free, but Agent and model conditions remain separate. Our Features page documents Full APK plugin support with setup and permissions; Play Lite cannot install APK plugins. Processing is not guaranteed wholly offline.

This workflow uses native saved recordings, not arbitrary external-audio or PDF import or a bridge to another interview service. Before analysis, confirm that every selected label refers to the intended session and that duplicate recordings are not being mistaken for additional interviews.

Read Three Clearly Fictional Interview Excerpts

The following three excerpts are entirely fictional teaching material. They are not participant data, a performed study, or actual FoneClaw output. Their purpose is to show how different preferences should remain distinguishable during analysis.

A possible paraphrase of fictional I01 is that the date is sometimes lost during a manual transfer. That wording is an analyst's restatement, not another quote. The excerpt does not explain which tool caused the loss or establish that automatic saving would solve it.

Fictional I02 expresses a desire to review a suggestion before saving. It does not state that every suggestion should be accepted. Fictional I03 explicitly rejects an additional reminder and instead asks for later retrieval of the original note.

Possible themes are date preservation, confirmation before saving, and source retrieval. A broader idea such as control or traceability is an analyst hypothesis to investigate, not a phrase all three fictional speakers used.

Keep the excerpts separate as you move into the matrix. The shared topic of reminders does not make the preferences identical. In particular, fictional I03 prevents a defensible summary from becoming a claim that all three want automatic reminders.

Build a Quote-to-Theme Evidence Matrix

Use the fictional excerpts to build a working matrix. Keep cells concise, but do not remove the distinction between evidence and interpretation merely to make the table shorter.

For actual interviews, add the saved recording name and a passage position you manually note while reviewing. The context above belongs to the fictional exercise; it is not an automatically generated timestamp or linked audio citation.

Check every quoted phrase against the transcript and audio, including nearby qualifications. If a proposed quote cannot be located, mark it missing or unverified. Do not generate plausible speech to fill an empty cell.

A translated restatement is not an exact original-language quote. Preserve the original wording when permitted and label any translation or paraphrase separately. Any conclusion broader than the checked passage belongs in the interpretation column, where readers can assess the inferential step.

Request Analysis Within the Selected Sources

Use one request that names your research question, selected labels, and required evidence boundaries. Keep unrelated recordings outside the selected collection context.

Editable request: Analyze only the selected interview recordings [labels] for [narrow research question]. Propose possible themes. For each theme, identify the correct recording label and provide exact source wording only when available. Keep paraphrases and analyst interpretations separate from quotes. Include conflicting cases, unclear passages, and questions for follow-up. Do not invent participant names, quotes, owners, deadlines, or product commitments. If a source or attribution is uncertain, mark it for review rather than filling the gap.

Then review each row, not just the overall theme list. Open the relevant recording, inspect the transcript, and listen to enough surrounding speech to understand the participant's meaning. Check whether a sentence answers the researcher's prompt, describes a past event, or imagines a future feature.

For the fictional exercise, a response that groups I03 under automatic-reminder demand needs correction. The objection is not decorative detail; it changes what the selected sources can support.

Record corrections in your working research note. If another answer is needed, first verify the intended sources and existing transcripts, then make a fresh bounded request. Do not assume that a note correction creates native version history or automatically refreshes every earlier answer.

Keep an unresolved source attribution unresolved. Asking for more confident wording does not establish which recording contains a statement or whether the participant said it.

Keep Interpretations and Contradicting Cases Visible

A theme is a way to organize evidence, not a vote count. Several mentions by one participant are not several independent users. The number of recordings selected also does not establish how common a preference is in a wider market.

In this fictional three-case exercise, the claim “everyone wants reminders saved automatically” is unsupported. It contradicts I02's request for review and I03's rejection of another reminder.

A bounded interpretation would be: within these fictional excerpts, one account describes date loss during copying, another asks for confirmation before saving, and another prioritizes finding the original note. Control and traceability are possible shared concerns, but the preferred response differs.

A useful next question is therefore: “When would you want a suggested reminder, and when would access to the original note be more useful?” That is a proposed research question, not a committed feature requirement.

Retain negative cases—evidence that challenges a proposed interpretation—in the same matrix as supporting cases. Do not place them in a separate appendix where the main finding appears unanimous.

For real research, also inspect whether the interviewer introduced the feature idea, whether the participant described actual behavior, and whether relevant context is missing. Silence about a feature is not agreement, and an unclear passage is not a negative response.

Keep the participant's words, your explanation of those words, and any proposed product response as separate layers. A feature wish can inform investigation without becoming an approved requirement, a causal explanation, or a claim about commercial potential.

Resolve Source Failures and Choose One Follow-Up

Before circulating the matrix, check the failures that can make a coherent analysis misleading. Return to the actual selected sources rather than repairing uncertainty with stronger prose.

Deliver a checked matrix, an explicit unknowns list, and one chosen research follow-up. Keep recordings, transcripts, analyst notes, and later confirmations distinct so a reader can see what supports each claim.

If you want to save that follow-up as a personal To-Do, make a separate supported request. Specify the real date when one is intended; otherwise request an Unscheduled entry and verify that no deadline was invented. Android AI To-Do Without a Due Date: Save It Safely explains that check.

Inspect the actual saved record and destination. The analysis itself does not automatically assign tasks, write to a CRM, contact participants, share research, or establish an export or native research dashboard.

Finish when the evidence is traceable and the remaining uncertainty has a useful next action. A smaller supported finding is more valuable than a broad conclusion that loses the original voices or their disagreement.