Recordings and transcripts

Design recording around the evidence your team needs. A caller’s response, voicemail and a recording of both sides are different requirements. Identify the capture source and the call that produces it before building a review process.

In a Bird sequence, configure recording through the steps available in the sequence builder. A recording window captures received audio from its source leg. A bridged call can produce separate recordings for each participant; do not assume one recording contains a mixed conversation. Ordinary SIP trunking does not automatically create a recording.

Define the recording requirement

1. State the purpose of capture and who needs to review it.
2. Identify the audio source. Check whether the connected application records received audio, generated prompts or both. A PBX or contact-center recording has its own source and permissions.
3. Include the notice and permission process appropriate to the conversation.
4. Define the capture window, access and retention requirements. Stop capture before sensitive input when the workflow requires it; hiding keypad telemetry does not remove speech from audio.
5. Decide whether the workflow needs transcription. A successful audio capture and a successful transcript are separate results.

Keep the source system explicit if your PBX or conversational runtime records the call. Moving telephone connectivity does not move its existing recording archive.

Prepare a controlled capture test

Use a number your team controls. Test the intended start, stop and hangup behavior, including a short call that ends before the normal capture window. Confirm which participant is audible and whether the expected audio was retained.

Keep the call, leg and recording identities with the result. A call can group related legs. Identify the leg that produced the recording. The leg log provides the telephone outcome.

Open Voice > Calls, then select the call to inspect its captured artifacts. To retrieve them through the API, list the call's recordings with the call_id from one of its legs. Each entry in recordings carries the leg_id and id that download its audio. The API returns each audio stream as its own file and does not mix streams. Entries with the same media_id refer to the same audio. The call transcript returns the cues of the call's transcribed streams in spoken order as JSON.

The audio operation answers 302 with a storage URL that is valid for 15 minutes. Request that URL without the Authorization header. Treat the URL as a secret: anyone holding it can hear the recording.

Distinguish processing states

Check audio availability and transcript processing independently. A pending transcript, failed transcription and unavailable audio require different responses. Empty text does not establish that the participant said nothing.

A call can end before processing finishes. Preserve the final telephone outcome while checking the artifact’s later state. If retained audio expires, retrying playback cannot restore it.

In the recording list, each stream in media has a state. While the state is open, capture is still running and the audio download answers 425: wait for the Retry-After interval, then request it again. A failed stream has no audio, and its download answers 404. Recordings and transcripts arrive separately after capture, so a recent call can list a stream before the recording entry that downloads it, and its transcript can still be missing.

truncated: true means a collection reached its limit of 500 entries. truncated: false does not establish that later recordings or transcripts have arrived. In the call transcript, complete: false means the transcription of a stream failed or the result was cut at its limit. complete: true does not establish that every stream was transcribed.

When a transcript is ambiguous, review the source audio where available. Names, dates and numbers can be misrecognized. Confirm important actions against the system that executed them.

Review access and retention

Use the permissions of the system that owns the artifact. Share a controlled call or recording reference with an authorized colleague rather than copying audio into unrelated records. Treat temporary playback URLs as temporary access, not a durable archive reference.

Bird keeps recordings and transcripts for 30 days from the start of the recorded audio. After that, a call's recordings and transcript no longer include them, and the audio download returns 404. Transcription does not restart the retention window. An empty result does not establish that nothing was ever recorded. For recordings stored by your PBX or conversational runtime, check that system's retention policy separately.

Confirm the business outcome

Keep source audio, transcript, generated summary and business result distinct. A summary can help a reviewer find an issue; the booking, payment or support system confirms whether an action completed.

Record missing or partial evidence in the review. A completed call does not establish that a recording exists, and a transcript alone does not prove that every participant was captured.

Next steps

• Call sequences
• Call log
• Voice events