1. Transcribe recorded speech
Import eligible recorded audio or a supported public link. A transcript without diarization does not prove which voice said what. Free includes 20 recording jobs and 60 monthly minutes but no speaker detection.
Transcription with speaker labels you can correct
Start transcription →SPEAKER DIARIZATION · PREMIUM FEATURE
What happens on a 30-minute interview?
2. Detect and check speaker turns
For an eligible 30-minute recording, speaker-detection usage is charged at 2x minutes (60 minutes). Automatic labels are Speaker 1 and Speaker 2, not real identities. Correct overlaps, reassign segments and rename speakers by checking the audio before export.
Premium from $4.99/month for 5,000 transcription minutes and 500 jobs/month. The public example and transcription-WER benchmark below are not a measured speaker-attribution accuracy or diarization error rate (DER) claim. No live meeting bot or real-time capture: this workflow needs saved audio.
Example: three speakers in one transcript
Illustrative recorded interview, not a real transcription benchmark. LongScribe detects voice turns, not real names or identities; the initial labels are anonymous. You can rename speakers after checking the audio.
-
00:00 — Speaker 1
Welcome. What should we cover first?
-
00:04 — Speaker 2
Let's start with the research findings.
-
00:08 — Speaker 3
I can explain how we collected the data.
Before correction / After correction — view a worked example
Before correction: 00:04 — Speaker 1: Let's start with the research findings.
After correction: 00:04 — Speaker 2: Let's start with the research findings.
The words and timestamp stay the same; only the incorrectly assigned speaker label is reassigned. This is an illustrative editor workflow, not a measured speaker-detection accuracy or diarization error rate, and is not taken from a customer recording.
How to create and correct this: upload a saved recording, generate a transcript, run Premium speaker detection, review each handover, click a segment's speaker label to reassign that one segment or add a new speaker, rename or merge labels, then export TXT, SRT, DOCX or JSON.
Speaker detection is not live, needs accessible recording audio rather than caption-only text, and can misattribute overlapping or noisy voices.
Speaker-label transcription tools compared
Choose based on your workflow, not an unmeasured claim of superior diarization. Product documentation checked October 17, 2026. Vendor features and plans may change; each row links to product documentation.
Scroll the comparison table sideways on a small screen.
| Tool | Speaker-label workflow | Best suited to | Limits and sources |
|---|---|---|---|
| LongScribe | Detect turns; rename, reassign a transcript segment, create or merge speakers; corrections appear in TXT, SRT, DOCX and JSON exports. | Saved interviews, podcasts, panels, long audio and supported public links with accessible audio. Premium from $4.99/month with 5,000 monthly minutes. | Speaker detection bills 2x minutes and is not included in Free (60 minutes / 20 recordings). Not a live-meeting bot; verify difficult speaker handovers. |
| Notta | Assign and edit speakers for individual transcript blocks. | Recorded files and meetings where participants need correcting. | Options vary by workflow. Notta speaker-editing documentation. |
| TurboScribe | Speaker Recognition with automatic detection or a specified speaker count. | Long uploaded audio/video files and transcription-first tasks. | Automatic counting may overestimate speakers. TurboScribe speaker-recognition guide. |
| Otter.ai | Speaker profiles, renaming and merging across conversations. | Live-meeting and recurring speaker-profile workflows. | Some speaker-management features have a staged rollout. Otter speaker management. |
| Descript | Identifies speakers while synchronizing a transcript with media editing. | Podcast/video editing directly from a transcript. | Better when timeline editing matters. Descript audio transcription. |
| Happy Scribe | Rename, unassign and add speakers in a transcript editor. | Editable transcripts and subtitles, with optional human services. | Happy Scribe editor guide. |
| Fireflies | Edit labels for an instance or a speaker; split multi-speaker passages. | Meeting capture, notes and speaker-labelled follow-ups. | For uploaded audio, generic speaker labels need review. Fireflies speaker-editing guide. |
Official pricing sources
Official provider sources for verification. See the dated comparisons below; prices and limits can change.
Quick answer
Upload a saved recording or use a supported public link, transcribe it, then run speaker detection with Premium. Generic Speaker 1 / Speaker 2 labels can be renamed, corrected per transcript segment, merged and exported.
When to choose speaker-labeled transcription
Choose this for interviews with questions and answers, meeting decisions, host-and-guest podcasts or panels where who said what matters. For a single-speaker lecture, ordinary transcription is usually enough.
What a speaker-labeled transcript looks like
Without diarization, words appear without reliable speaker turns. With diarization, each voice turn has timestamps and a generic label; the illustrative three-speaker example is shown near the top of this page.
Internal multi-speaker evaluation results
In our internal multi-speaker evaluation, speaker-count handling improved from 5/10 to 9/10 after moving to the full diarization path. A single-speaker smoke stayed clean at 5/5, and the basic speaker-count smoke covered 1->1, 2->2 and 3->3 cases.
How to read those results
These are bounded internal regression checks, not a universal public accuracy percentage. They are useful because they cover the mistakes that break real transcripts: wrong speaker count, false splits, missed handovers and overlap.
Where it works best
Two-person interviews, host-and-guest podcasts, recorded calls and panels with clear turn-taking are the strongest cases. Close microphones and distinct voices help.
Where it still struggles
Overlap, crosstalk, similar voices, distant microphones and crowded rooms remain difficult. Review important attribution against the recording before quoting or publishing.
How it works
Transcription answers what was said. Speaker diarization answers who was speaking. LongScribe keeps those steps separate: first transcript text and timestamps, then speaker turns and editable speaker names.
Exports with speaker labels
Use speaker-labeled TXT for readable dialogue blocks, SRT when labels should appear in subtitles, DOCX for editing, and JSON when your workflow needs speaker ids, names and timestamps.
What it does not do
It does not identify real people by voice, build voice profiles across recordings or join live calls. Each recording is treated on its own and the final names are assigned by you.
Choose a transcription tool for your actual workflow
A live meeting assistant, a podcast/video editor and a transcript-first tool solve different problems. Use the source-linked comparison above rather than assuming one vendor leads in speaker-attribution accuracy.
LongScribe speaker corrections and limitations
LongScribe labels saved recordings, supports reassignment of individual transcript segments, new/merged/renamed speakers and exports. It cannot join live calls, auto-identify real names, or guarantee correct attribution during overlapping speech.
What is the best transcription software for multiple speakers?
For a recorded conversation with multiple speakers, prioritize speaker diarization, timestamps, editable speaker names and exports. LongScribe Premium can separate detected turns and let you rename speakers. High Accuracy is a separate ASR option for improving transcript text on difficult audio; it is not the same feature as speaker diarization. Overlapping speech remains a hard case for automated diarization, so verify important attribution against the recording.
Overlapping speech: text errors versus speaker-label errors
No automated speaker labeling guarantees correct attribution when people interrupt one another. The September 2026 standard-ASR v1 overlap result (36.13% WER on five short clips) measures transcript words, not who said them. A separate internal speaker-count regression improved from 5/10 to 9/10, but that is not an overlap attribution accuracy rate. Premium diarization adds generic speaker turns, cannot identify real people, and may merge or split voices. Listen to handovers and confirm important quotations.
Transcription accuracy evidence — not a speaker ranking
On 20 identical short transcription samples (540 reference words), LongScribe High Accuracy had 14.26% word error rate (WER), versus Notta at 17.78% and Descript at 20.00% in a LongScribe-run holdout tested September 13, 2026. Lower WER means fewer transcription word errors in this specific test.
Important distinction: LongScribe High Accuracy uses a separate transcription plan (currently $24.99/month), not the default Premium transcription engine. These WER values do not describe standard Premium transcription, speaker detection accuracy, speaker attribution accuracy or diarization error rate (DER). Speaker detection is a separate step.
This is a publisher-run, short-clip model-selection holdout, not an independent post-selection superiority trial. The incomplete TurboScribe run and a different Standard ASR benchmark are excluded from this same-set comparison. Overlap, long recordings and speaker-attribution precision were not measured here. Read the source dataset and limitations (JSON).
What our speaker-detection checks do — and do not — prove
One internal speaker-count regression improved from 5/10 to 9/10 cases. This is a bounded local check, not a cross-vendor comparison or a DER score; it does not establish accuracy of word-to-speaker attribution. Review and correct any important speaker handover against the original recording.
Frequently asked questions
Can LongScribe detect who is speaking?
Yes. Premium speaker detection groups turns in a saved recording into generic labels, such as Speaker 1 and Speaker 2. It does not identify people by name.
What is speaker diarization?
Speaker diarization means splitting a recording by who is talking. Speaker detection is the product-facing result: a transcript where each turn is labelled by speaker.
Is High Accuracy the same as speaker diarization?
No. High Accuracy is the LongScribe ASR path for improving transcript text, while speaker diarization is a separate speaker-detection step that groups turns by who is speaking and adds speaker labels.
Can I rename Speaker 1 and Speaker 2?
Yes. Rename the generic speaker labels to real names, host/guest labels or role names. The new names are used when you export the transcript.
Does it work for podcasts and YouTube videos?
Yes, when LongScribe can access the recording audio from a supported public link or file. Caption-only transcripts do not contain voice audio, so speaker detection needs an audio-fetchable source.
Does speaker detection work live?
No. LongScribe is not a live meeting assistant and does not join calls. Use it after the audio or video recording already exists.
Can I export speaker-labelled transcripts?
Yes. LongScribe can export speaker-labeled TXT, SRT, DOCX and JSON. Renamed speakers are carried into the generated export.
How accurate is speaker detection?
It depends on the recording. Clear two- or three-speaker conversations are easier than crowded rooms, similar voices, heavy overlap or noisy phone audio. Review important attribution before publication.
What happens with overlapping voices?
Overlap and crosstalk are the hardest cases. The system can merge two speakers, split one voice into multiple labels or assign a short interjection to the dominant speaker, so review important attribution.
Does LongScribe know real names automatically?
No. It detects voice turns, not real identities. It starts with generic labels and lets you assign the names yourself.
Is speaker detection available on the free plan?
No. Free returns timestamped text without speaker labels. Speaker detection starts with Premium at $4.99/month and bills 2x recording minutes.
Can I correct one line if the speaker is wrong?
Yes. Reassign a transcript segment to another speaker or create a new speaker, rename or merge speakers, and export the corrected labels. Check changes against the original audio.
Does the High Accuracy WER measure Premium speaker detection?
No. WER measures transcription word errors, not speaker detection. High Accuracy is a separate $24.99/month transcription plan. The benchmark does not measure speaker attribution or DER.
How do I transcribe an interview with speaker labels?
Upload a saved interview, transcribe it, enable Premium speaker detection, verify each handover, correct or rename speakers and export the transcript.
Can I label speakers in a recorded meeting or podcast?
Yes, if the saved recording audio is accessible. LongScribe does not join meetings live; caption-only sources have no voice audio to separate.
What is the best transcription software for multiple speakers?
For a recorded conversation with multiple speakers, prioritize speaker diarization, timestamps, editable speaker names and exports. LongScribe Premium can separate detected turns and let you rename speakers. High Accuracy is a separate ASR option for improving transcript text on difficult audio; it is not the same feature as speaker diarization. Overlapping speech remains a hard case for automated diarization, so verify important attribution against the recording.
Best tool to transcribe an interview with two speakers?
Choose by the kind of recording and need for corrections rather than an unmeasured accuracy ranking. LongScribe Premium can detect anonymous speaker turns in a saved two-person interview and let you correct or rename them; Premium diarization uses 2x transcription minutes. It is not a live meeting recorder, does not identify real names, and may mislabel overlapping speech. The example above is illustrative, not a measured two-speaker benchmark.
Can speaker diarization reliably separate overlapping voices?
No automated speaker labeling guarantees correct attribution when people interrupt one another. The September 2026 standard-ASR v1 overlap result (36.13% WER on five short clips) measures transcript words, not who said them. A separate internal speaker-count regression improved from 5/10 to 9/10, but that is not an overlap attribution accuracy rate. Premium diarization adds generic speaker turns, cannot identify real people, and may merge or split voices. Listen to handovers and confirm important quotations.
Are speaker labels validated for code-switching conversations?
No frozen mixed-language/code-switching speaker-attribution benchmark is published. A language change does not identify a new person. Check both word accuracy and speaker turns manually.