A WhatsApp Summary Without the Voice Messages Is Wrong

A WhatsApp summary without the voice messages is not a short version of the conversation. It is a different conversation. The exported chat log writes `<Media omitted>` where every voice message used to be, so a tool that reads the text alone summarises whatever was typed and presents it as the whole — confidently, and with no warning that anything is missing. This page is about closing that gap: transcribing the audio and putting it back where it belongs in the timeline. The voice-to-text tool is built around exactly that pipeline.

The correct workflow is:

1. Export the chat as a `.zip` with media.
2. Transcribe every voice message.
3. Merge transcripts into the chat timeline at the original timestamps.
4. Run analysis on the combined stream and extract decisions, action items, and open questions.

This page is the working playbook for that workflow at scale, including the parts most guides skip, what `.opus` actually is, why the merge step matters more than the transcription step, and how to keep group chats useful when half the participants only ever send 30-second voice memos.

The voice-to-text tool is built around this exact pipeline.

What WhatsApp voice messages actually are

WhatsApp records voice messages with the Opus audio codec inside an OGG container, exported as `.opus` files. Older iOS exports occasionally use `.m4a` (AAC inside an MP4 container).

Technical specifics:

• Codec: Opus in voice-over-IP mode.
• Bitrate: roughly 16 kbps.
• Channels: mono.
• Sample rate: 16 kHz.
• Container: OGG (`.opus`) or MP4 (`.m4a`).

Two consequences:

1. Compression is aggressive. Opus at 16 kbps preserves intelligibility but strips most of the harmonic detail above 8 kHz. Sibilants and unvoiced stops are the first things to degrade on a poor connection.
2. The sample rate is fixed at the codec's own. Nothing has to be resampled before transcription, but there is no audio above 8 kHz left to recover either, which sets a hard limit on what any speech-to-text can hear.

If you export with media, the `.zip` includes the audio files alongside `_chat.txt`. If you export without media, the audio files are missing entirely and the chat log shows `<attached: ...opus>` placeholder lines or `audio omitted` text where the voice messages used to be.

Practical takeaway: no media, no audio transcription. Re-export with media if you missed it the first time. If you only need the audio side and not the full chat, the .opus to readable text tool handles the same files in isolation.

Step 1: Export the chat with media

iPhone

1. Open the chat.
2. Tap the contact or group name at the top.
3. Scroll to Export Chat.
4. Choose Attach Media.
5. Tap Save to Files and pick a location. The `.zip` stays on the phone and you can find it again in the Files app.

Android

1. Open the chat.
2. Tap the menu (three dots, top right).
3. Tap More.
4. Tap Export chat.
5. Choose Include media.
6. WhatsApp hands the finished `.zip` to the share sheet, once. Whatever you tap there is the only place a copy will exist: dismiss the sheet and the file is in no folder on the phone, and the export has to be redone.

On Samsung (One UI) the sheet has a Save as file entry, sometimes behind More, which writes the `.zip` to Downloads. On every other Android there is no such entry, so send it to Google Drive. When you upload later, the Android file picker lists Google Drive in its side menu, so you pick the `.zip` straight from there with no download.

Step 2: Verify the export contains voice messages

Inside the `.zip`, you should see:

• A chat text file (often `_chat.txt`, sometimes `WhatsApp Chat - <name>.txt`).
• Multiple `.opus` or `.m4a` audio files (one per voice message).
• Image, video, and other media files if any were sent.

If you do not see `.opus` or `.m4a` files, the export was made without media. Re-export.

If you see them but they are all very small (under 1 KB), the export hit a media-size cap and the audio is corrupted. Re-export with a smaller date range.

Step 3: Bulk transcription strategy (the only one that scales)

Transcribing voice messages one by one is a waste of time. A scalable pipeline does this automatically:

• Parse the chat log and detect every voice message reference (`<attached: ...opus>` lines).
• Match each reference to the actual `.opus` or `.m4a` file inside the `.zip`.
• Decode the audio and split any clip that exceeds the transcription endpoint's size limit.
• Transcribe each clip through a speech-to-text API.
• Return per-clip results: text, language, duration, timestamps inside the clip.
• Merge transcripts into the conversation timeline at the original send timestamps.

That last step is the difference between "a pile of audio transcripts" and "a usable recap". Most tools that advertise WhatsApp voice transcription stop at step five and leave the merge as a manual exercise.

Step 4: Merge transcripts into the timeline

A correctly merged transcript looks like a normal message in the conversation timeline:

• Sender: Alex.
• Type: audio.
• Timestamp: 14:32:11 on 27 January 2026 (original send time).
• Transcript: "Ok, we will ship Friday. John owns the landing page. I will handle billing."

With this structure, downstream analysis can correctly extract:

• Decisions: ship Friday.
• Owners: John for landing page.
• Action items: billing tasks (owner: speaker).
• Open questions: anything unresolved in the transcript.

Without timeline merge, the AI sees the chat log without audio content and the audio transcripts as a separate disconnected stream. The recap then misses commitments made only in audio, which in many work chats is the majority of substantive content.

This is the most common failure mode of generic transcription tools paired with general-purpose summarisers.

Step 5: Turn transcripts into real outputs

Once audio is merged into the timeline, the choice of analysis goal shapes what you get:

Meeting Recap

• Context and purpose.
• Agenda topics in order.
• Decisions made (with the deciding speaker and timestamp).
• Action items (owner, deadline if mentioned, current state).
• Open questions.
• Suggested follow-ups.

Best for project standups, sprint planning, retros conducted in WhatsApp. The same output reads cleanly as meeting minutes from a WhatsApp chat when you need a sharable artefact.

Action Items only

• Task list.
• Owner per task.
• Deadline or "no deadline mentioned".
• Blockers.

Best when you only need a current commitments list and the broader context is not needed.

Conflict Resolution

• Root cause.
• Each side's perspective.
• Misunderstandings.
• Resolution status.
• Next steps.

Best for arguments and disagreements that played out in audio. Voice tone often matters here, but the transcript captures the content even if it loses the tone.

Decisions

• Decision text.
• Who decided.
• Supporting context.
• Dissent (if any).
• Date and timestamp.

Best for project history audits or when you need a defensible record of what was agreed and when.

Relationship Insights

• Tone arc over time.
• Recurring topics.
• Communication patterns.

Best for personal or partnership chats where the value is in the longitudinal view rather than specific commitments. The full output shape is documented under relationship insights from WhatsApp history.

Accuracy tips, simple and high-impact

Transcription quality follows audio quality. Practical levers:

• Distance: phone 10–20 cm from the mouth. Closer than that introduces breath and plosive noise; further than that picks up room reverb.
• Pace: moderate, not rushed. The model handles natural conversational speech well; rushed speech compounds errors at chunk boundaries.
• Environment: indoors beats outdoors. Stationary beats walking. Silent room beats music or TV in background.
• Names and numbers: state them deliberately. If a name or invoice number matters, say it twice ("invoice four-seven-two-nine, four seven two nine").
• One language per clip: code-switching mid-sentence is the hardest case for the model. Switching across a sentence break is fine.

For more detail on what accuracy to expect under different conditions, see the voice message accuracy reference.

Common problems and fixes

My export is missing voice messages

Re-export with Include media (Android) or Attach Media (iPhone). Without media, the audio files are not in the `.zip` at all.

My `.zip` is too big to upload

Start with a smaller timeframe. If you only need "what happened this week," do not export three years of media. WhatsApp also caps exports at 10,000 messages when media is included; for very long chats, run two exports, one without media for full historical coverage, one with media for the recent period that contains the voice messages you actually need.

The tool transcribed audio but the recap is still generic

Almost always means the transcripts were not merged into the conversation timeline before analysis. Audio transcripts as a separate document do not carry conversational context, so the analysis cannot reason about who said what and when. ThreadRecap performs the merge automatically; if you are using a different tool, this step is usually missing.

Group chats are noisy

The transcript got names wrong

Expected behaviour for machine transcription, proper nouns are the most common error category. Spot-check names against the original audio using the inline player (every transcribed clip in ThreadRecap has a player at the message position). Names that appear repeatedly in the chat tend to converge on the right spelling because the model has more context to anchor on.

Privacy basics for voice messages

Voice messages can include identity cues, names, locations, and confidential details. The minimum a serious tool should provide:

• Preview of what will be processed before upload.
• Selective upload: chat text and audio sent to servers; photo, video, and document files never uploaded, and from a video only the audio track extracted in the browser.
• Encrypted account storage for chat text, voice message audio, and processed recaps, with explicit user control over deletion.
• Clear retention policy in writing.
• No model training on user-uploaded content.

ThreadRecap parses `.zip` files locally in the browser, never uploads photos, videos, or documents, stores chat text and voice message audio encrypted in your account alongside processed recaps, and gives you deletion control through the dashboard at any time. Review the privacy policy for retention specifics before uploading sensitive content. If you select a video, your browser extracts and sends only its audio track for transcription. The video and its visual content stay on your device.

Quick reference

Can I transcribe WhatsApp voice messages to text for free?

What file format are WhatsApp voice messages?

`.opus` (Opus codec, OGG container) is the default. `.m4a` (AAC, MP4 container) appears in older iOS exports. Both inside the export `.zip` when media is included.

Do I need media export for transcription?

Yes. No media in the export means no audio files to transcribe.

What is the best end result to aim for?

A searchable timeline where voice messages are merged back into the conversation, plus a structured output such as decisions and action items with owners and deadlines. The transcript on its own is much less useful than the same transcript inside the conversational context.

Run the workflow

Export your WhatsApp chat with media, upload the `.zip`, and let ThreadRecap transcribe the voice messages in one pass. Start with the first summary, then ask follow-up questions about decisions, action items, or another topic you want to review.