ElevenLabs Scribe
Fast batch transcription with rich metadata, strongest on jargon and weaker on overlap and code-switching.
Strong batch STT API with rich metadata and clear caveats
- You need batch speech-to-text with word timestamps, confidence values, and speaker labels in the payload.
- You need strong accuracy on jargon-heavy English narration.
- You need fast batch turnaround on long recordings and can work with about $0.2202/audio-hour.
- You need validated speaker attribution correctness in overlapping conversations.
Our take
ElevenLabs Scribe is a strong batch speech-to-text API for developers: every scored run returned word-level timestamps, confidence values, and speaker labels, and it was outstanding on the jargon-heavy sample. The tradeoff is that crosstalk produced heavy insertions and over-segmentation, while the bilingual run lost Spanish tokens and the trace recorded an HTTP 401, so it reads as a fast English-first transcription backend with clear limits on overlap and code-switching.
In-Depth Review
Our detailed analysis of ElevenLabs Scribe — features, performance, and real-world testing.
Feature-by-Feature Breakdown
Batch Speech-to-Text TranscriptionMixed▾
Feature tested: Batch Speech-to-Text Transcription
Result: Partial
Verdict: Mixed
Expected behavior: The API transcribes longer or batch audio in a single request, including the medical narration, crosstalk, bilingual sample, and 18–36 minute recordings. The sampled cards use those runs as evidence for overall transcription throughput and quality across hard-audio variants.
Test case: Audio file → Image
Input type: Audio file
Input used: Input artifact (Audio file): INPUT: Overlapping speech / crosstalk audio (crosstalk.wav; 65.39 MB; 2142.709s; four-person meeting overlap). — crosstalk.wav
Observed output: Output artifact (Image): Crosstalk transcript detail: the engine skipped a long overlapping span, and the scored run landed at WER 26.67% with 856 substitutions, 782 deletions, and 383 insertions. — 04-transcript-detail.png
Input artifact: Input artifact (Audio file): INPUT: Overlapping speech / crosstalk audio (crosstalk.wav; 65.39 MB; 2142.709s; four-person meeting overlap). — crosstalk.wav
Output artifact: Output artifact (Image): Crosstalk transcript detail: the engine skipped a long overlapping span, and the scored run landed at WER 26.67% with 856 substitutions, 782 deletions, and 383 insertions. — 04-transcript-detail.png
What changed: Audio file transformed into Image
Test case: Audio file → Image
Input type: Audio file
Input used: Input artifact (Audio file): INPUT: Medical jargon narration (medical_terms.mp3; 8.58 MB; 1123.944s; Gray's Anatomy / LibriVox). — medical_terms.mp3
Observed output: Output artifact (Image): Medical-jargon transcript detail: the engine stayed close to the reference, with WER 3.01% and perfect jargon recall on the scored terms. — 04-transcript-detail-2.png
Input artifact: Input artifact (Audio file): INPUT: Medical jargon narration (medical_terms.mp3; 8.58 MB; 1123.944s; Gray's Anatomy / LibriVox). — medical_terms.mp3
Output artifact: Output artifact (Image): Medical-jargon transcript detail: the engine stayed close to the reference, with WER 3.01% and perfect jargon recall on the scored terms. — 04-transcript-detail-2.png
What changed: Audio file transformed into Image
Test case: Audio file → Image
Input type: Audio file
Input used: Input artifact (Audio file): INPUT: Bilingual code-switching audio (mix_language.mp3; 22.19 MB; 1938.495s; Spanish-English conversation). — mix_language.mp3
Observed output: Output artifact (Image): Bilingual transcript detail: the engine dropped Spanish token 'ahora' and rewrote it as 'oh now', with WER 28.57% and Spanish recall 57.5%. — 04-transcript-detail-3.png
Input artifact: Input artifact (Audio file): INPUT: Bilingual code-switching audio (mix_language.mp3; 22.19 MB; 1938.495s; Spanish-English conversation). — mix_language.mp3
Output artifact: Output artifact (Image): Bilingual transcript detail: the engine dropped Spanish token 'ahora' and rewrote it as 'oh now', with WER 28.57% and Spanish recall 57.5%. — 04-transcript-detail-3.png
What changed: Audio file transformed into Image
Test case: Audio file → Image
Input type: Audio file
Input used: Input artifact (Audio file): INPUT: input-1 — Overlapping Speech / Crosstalk (crosstalk.wav, 35:43, 65.39 MB, overlapping meeting audio with 4 participants). — crosstalk.wav
Observed output: Output artifact (Image): Inline JSON transcript returned for the crosstalk sample; the benchmark later scored it at 26.67% WER with 41.76s latency. — 02-response-raw.png
Input artifact: Input artifact (Audio file): INPUT: input-1 — Overlapping Speech / Crosstalk (crosstalk.wav, 35:43, 65.39 MB, overlapping meeting audio with 4 participants). — crosstalk.wav
Output artifact: Output artifact (Image): Inline JSON transcript returned for the crosstalk sample; the benchmark later scored it at 26.67% WER with 41.76s latency. — 02-response-raw.png
What changed: Audio file transformed into Image
Test case: Audio file → Image
Input type: Audio file
Input used: Input artifact (Audio file): INPUT: input-2 — Medical Jargon (medical_terms.mp3, 18:44, 8.58 MB, single-speaker Gray's Anatomy narration). — medical_terms.mp3
Observed output: Output artifact (Image): Inline JSON transcript returned for the medical-jargon sample; the benchmark later scored it at 3.01% WER with 16.33s latency. — 02-response-raw-2.png
Input artifact: Input artifact (Audio file): INPUT: input-2 — Medical Jargon (medical_terms.mp3, 18:44, 8.58 MB, single-speaker Gray's Anatomy narration). — medical_terms.mp3
Output artifact: Output artifact (Image): Inline JSON transcript returned for the medical-jargon sample; the benchmark later scored it at 3.01% WER with 16.33s latency. — 02-response-raw-2.png
What changed: Audio file transformed into Image
Test case: Audio file → Image
Input type: Audio file
Input used: Input artifact (Audio file): INPUT: input-3 — Bilingual Code-Switching (mix_language.mp3, 32:18, 22.19 MB, spontaneous Spanish-English conversation). — mix_language.mp3
Observed output: Output artifact (Image): The raw response identifies the sample as English with 0.967 language probability and includes speaker-labeled word timings, but Spanish coverage remains partial. — 02-response-raw-3.png
Input artifact: Input artifact (Audio file): INPUT: input-3 — Bilingual Code-Switching (mix_language.mp3, 32:18, 22.19 MB, spontaneous Spanish-English conversation). — mix_language.mp3
Output artifact: Output artifact (Image): The raw response identifies the sample as English with 0.967 language probability and includes speaker-labeled word timings, but Spanish coverage remains partial. — 02-response-raw-3.png
What changed: Audio file transformed into Image
Test case: Audio file → Text/code file
Input type: Audio file
Input used: Input artifact (Audio file): INPUT-1: Overlapping Speech / Crosstalk — crosstalk.wav (35:43, 65.4 MB, AMI EN2002a). — crosstalk.wav
Observed output: Output artifact (Text/code file): Returned a scored transcript for the crosstalk file, with 7,180 hypothesis words versus 7,579 reference words and WER 26.67%; the run completed as a batch transcription job rather than failing outright. — raw-response.json
Input artifact: Input artifact (Audio file): INPUT-1: Overlapping Speech / Crosstalk — crosstalk.wav (35:43, 65.4 MB, AMI EN2002a). — crosstalk.wav
Output artifact: Output artifact (Text/code file): Returned a scored transcript for the crosstalk file, with 7,180 hypothesis words versus 7,579 reference words and WER 26.67%; the run completed as a batch transcription job rather than failing outright. — raw-response.json
What changed: Audio file transformed into Text/code file
Test case: Audio file → Text/code file
Input type: Audio file
Input used: Input artifact (Audio file): INPUT-2: Medical Jargon — medical_terms.mp3 (18:44, 8.6 MB, Gray's Anatomy via LibriVox). — medical_terms.mp3
Observed output: Output artifact (Text/code file): Returned a scored transcript for the medical-jargon file, with 2,743 hypothesis words versus 2,728 reference words and WER 3.01%, which was the strongest result in this set. — raw-response-2.json
Input artifact: Input artifact (Audio file): INPUT-2: Medical Jargon — medical_terms.mp3 (18:44, 8.6 MB, Gray's Anatomy via LibriVox). — medical_terms.mp3
Output artifact: Output artifact (Text/code file): Returned a scored transcript for the medical-jargon file, with 2,743 hypothesis words versus 2,728 reference words and WER 3.01%, which was the strongest result in this set. — raw-response-2.json
What changed: Audio file transformed into Text/code file
Test case: Audio file → Text/code file
Input type: Audio file
Input used: Input artifact (Audio file): INPUT-3: Bilingual Code-Switching — mix_language.mp3 (32:18, 22.2 MB, Bangor Miami herring1). — mix_language.mp3
Observed output: Output artifact (Text/code file): Returned a scored transcript for the bilingual file, with 6,071 hypothesis words versus 6,517 reference words and WER 28.57%; the execution trace for this run also recorded an HTTP 401 error. — raw-response-3.json
Input artifact: Input artifact (Audio file): INPUT-3: Bilingual Code-Switching — mix_language.mp3 (32:18, 22.2 MB, Bangor Miami herring1). — mix_language.mp3
Output artifact: Output artifact (Text/code file): Returned a scored transcript for the bilingual file, with 6,071 hypothesis words versus 6,517 reference words and WER 28.57%; the execution trace for this run also recorded an HTTP 401 error. — raw-response-3.json
What changed: Audio file transformed into Text/code file
Test case: Text prompt → Image
Input type: Text prompt
Input used: Input artifact (Text prompt): INPUT
Observed output: Output artifact (Image): The crosstalk run finished in 41.76s wall clock at RTF 0.01949 with an estimated cost of $0.13106. — 03-terminal-metrics.png
Input artifact: Input artifact (Text prompt): INPUT
Output artifact: Output artifact (Image): The crosstalk run finished in 41.76s wall clock at RTF 0.01949 with an estimated cost of $0.13106. — 03-terminal-metrics.png
What changed: Text prompt transformed into Image
Test case: Text prompt → Image
Input type: Text prompt
Input used: Input artifact (Text prompt): INPUT
Observed output: Output artifact (Image): The medical-jargon run finished in 16.33s wall clock at RTF 0.01453 with an estimated cost of $0.06875. — 03-terminal-metrics-2.png
Input artifact: Input artifact (Text prompt): INPUT
Output artifact: Output artifact (Image): The medical-jargon run finished in 16.33s wall clock at RTF 0.01453 with an estimated cost of $0.06875. — 03-terminal-metrics-2.png
What changed: Text prompt transformed into Image
Test case: Text prompt → Image
Input type: Text prompt
Input used: Input artifact (Text prompt): INPUT
Observed output: Output artifact (Image): The bilingual run finished in 6.07s wall clock at RTF 0.00313 with an estimated cost of $0.11857, making it the fastest of the three. — 03-terminal-metrics-3.png
Input artifact: Input artifact (Text prompt): INPUT
Output artifact: Output artifact (Image): The bilingual run finished in 6.07s wall clock at RTF 0.00313 with an estimated cost of $0.11857, making it the fastest of the three. — 03-terminal-metrics-3.png
What changed: Text prompt transformed into Image
Why it matters / Conclusion: Best on the medical jargon sample, middling on crosstalk, and weaker on bilingual spans, so it is useful but not evenly strong across the hard-audio set.
The API transcribes longer or batch audio in a single request, including the medical narration, crosstalk, bilingual sample, and 18–36 minute recordings. The sampled cards use those runs as evidence for overall transcription throughput and quality across hard-audio variants.
Transcript Metadata▾
Feature tested: Transcript Metadata
Result: Passed
Expected behavior: The API returns structured transcript payloads with word-level timestamps, confidence values, speaker labels, and other annotation fields. The evidence is drawn from raw JSON transcript output and per-token metadata that was consistently present across the sampled runs.
Test case: Audio file → Image
Input type: Audio file
Input used: Input artifact (Audio file): INPUT: input-1 — Overlapping Speech / Crosstalk (crosstalk.wav, 35:43, 65.39 MB, overlapping meeting audio with 4 participants). — crosstalk.wav
Observed output: Output artifact (Image): The raw response includes word objects with `start`/`end` times and `logprob` confidence, and the run reported 14,506 timed tokens on this crosstalk sample. — 02-response-raw.png
Input artifact: Input artifact (Audio file): INPUT: input-1 — Overlapping Speech / Crosstalk (crosstalk.wav, 35:43, 65.39 MB, overlapping meeting audio with 4 participants). — crosstalk.wav
Output artifact: Output artifact (Image): The raw response includes word objects with `start`/`end` times and `logprob` confidence, and the run reported 14,506 timed tokens on this crosstalk sample. — 02-response-raw.png
What changed: Audio file transformed into Image
Test case: Audio file → Image
Input type: Audio file
Input used: Input artifact (Audio file): INPUT: input-2 — Medical Jargon (medical_terms.mp3, 18:44, 8.58 MB, single-speaker Gray's Anatomy narration). — medical_terms.mp3
Observed output: Output artifact (Image): The raw response includes word objects with `start`/`end` times and `logprob` confidence, and the run reported 5,448 timed tokens on the medical sample. — 02-response-raw-2.png
Input artifact: Input artifact (Audio file): INPUT: input-2 — Medical Jargon (medical_terms.mp3, 18:44, 8.58 MB, single-speaker Gray's Anatomy narration). — medical_terms.mp3
Output artifact: Output artifact (Image): The raw response includes word objects with `start`/`end` times and `logprob` confidence, and the run reported 5,448 timed tokens on the medical sample. — 02-response-raw-2.png
What changed: Audio file transformed into Image
Test case: Audio file → Image
Input type: Audio file
Input used: Input artifact (Audio file): INPUT: input-3 — Bilingual Code-Switching (mix_language.mp3, 32:18, 22.19 MB, spontaneous Spanish-English conversation). — mix_language.mp3
Observed output: Output artifact (Image): The raw response includes word objects with `start`/`end` times and `logprob` confidence, and the run reported 12,162 timed tokens on the bilingual sample. — 02-response-raw-3.png
Input artifact: Input artifact (Audio file): INPUT: input-3 — Bilingual Code-Switching (mix_language.mp3, 32:18, 22.19 MB, spontaneous Spanish-English conversation). — mix_language.mp3
Output artifact: Output artifact (Image): The raw response includes word objects with `start`/`end` times and `logprob` confidence, and the run reported 12,162 timed tokens on the bilingual sample. — 02-response-raw-3.png
What changed: Audio file transformed into Image
Why it matters / Conclusion: This metadata is consistently present and usable downstream on every scored run.
The API returns structured transcript payloads with word-level timestamps, confidence values, speaker labels, and other annotation fields. The evidence is drawn from raw JSON transcript output and per-token metadata that was consistently present across the sampled runs.
Speaker Diarization▾
Feature tested: Speaker Diarization
Result: Partial
Expected behavior: The API assigns speaker labels in transcript output and reports distinct-speaker counts for audio such as crosstalk, narration, and bilingual conversation. The sampled cards emphasize label presence in the transcript payload and downstream use for analytics or captions.
Test case: Audio file → Image
Input type: Audio file
Input used: Input artifact (Audio file): Input 1: Overlapping Speech / Crosstalk — four participants talking over one another. — crosstalk.wav
Observed output: Output artifact (Image): The raw response shows speaker labels in the JSON and reports 5 distinct speakers on a source with 4 participants. — 02-response-raw.png
Input artifact: Input artifact (Audio file): Input 1: Overlapping Speech / Crosstalk — four participants talking over one another. — crosstalk.wav
Output artifact: Output artifact (Image): The raw response shows speaker labels in the JSON and reports 5 distinct speakers on a source with 4 participants. — 02-response-raw.png
What changed: Audio file transformed into Image
Test case: Audio file → Image
Input type: Audio file
Input used: Input artifact (Audio file): INPUT-2: Medical Jargon — check diarization behavior on a single-speaker narration. — medical_terms.mp3
Observed output: Output artifact (Image): The medical-jargon run detected a single speaker label, which matches the single-reader nature of the audio. — 02-response-raw-2.png
Input artifact: Input artifact (Audio file): INPUT-2: Medical Jargon — check diarization behavior on a single-speaker narration. — medical_terms.mp3
Output artifact: Output artifact (Image): The medical-jargon run detected a single speaker label, which matches the single-reader nature of the audio. — 02-response-raw-2.png
What changed: Audio file transformed into Image
Test case: Audio file → Image
Input type: Audio file
Input used: Input artifact (Audio file): INPUT-3: Bilingual Code-Switching — check diarization behavior on mixed-language speech. — mix_language.mp3
Observed output: Output artifact (Image): The bilingual run detected 3 speaker labels, showing the payload exposes diarization even on mixed-language audio. — 02-response-raw-3.png
Input artifact: Input artifact (Audio file): INPUT-3: Bilingual Code-Switching — check diarization behavior on mixed-language speech. — mix_language.mp3
Output artifact: Output artifact (Image): The bilingual run detected 3 speaker labels, showing the payload exposes diarization even on mixed-language audio. — 02-response-raw-3.png
What changed: Audio file transformed into Image
Test case: Audio file → Image
Input type: Audio file
Input used: Input artifact (Audio file): INPUT-1: Overlapping Speech / Crosstalk — check diarization behavior on a 4-participant meeting. — crosstalk.wav
Observed output: Output artifact (Image): The crosstalk sample over-segmented the conversation: the engine detected 5 speaker labels on a file with 4 participants, so diarization presence is real but attribution quality is not proven. — 04-transcript-detail.png
Input artifact: Input artifact (Audio file): INPUT-1: Overlapping Speech / Crosstalk — check diarization behavior on a 4-participant meeting. — crosstalk.wav
Output artifact: Output artifact (Image): The crosstalk sample over-segmented the conversation: the engine detected 5 speaker labels on a file with 4 participants, so diarization presence is real but attribution quality is not proven. — 04-transcript-detail.png
What changed: Audio file transformed into Image
Why it matters / Conclusion: The labels exist, but the benchmark only measured label presence and count, not attribution correctness.
The API assigns speaker labels in transcript output and reports distinct-speaker counts for audio such as crosstalk, narration, and bilingual conversation. The sampled cards emphasize label presence in the transcript payload and downstream use for analytics or captions.
Code-Switching / Multilingual Transcription▾
Feature tested: Code-Switching / Multilingual Transcription
Result: Passed
Expected behavior: The API transcribes mixed-language audio without special configuration, as shown on the Spanish-English sample. The evidence shows successful mixed-language transcription with weaker Spanish recall than clean-English narration.
Test case: Audio file → Image
Input type: Audio file
Input used: Input artifact (Audio file): INPUT-3: Bilingual Code-Switching — Spanish-English conversation with spontaneous switching. — mix_language.mp3
Observed output: Output artifact (Image): The bilingual diagnostic shows the engine dropped the Spanish token 'ahora' and rewrote it as 'oh now', with Spanish recall reported at 57.5% (46 of 80 types). — 04-transcript-detail-3.png
Input artifact: Input artifact (Audio file): INPUT-3: Bilingual Code-Switching — Spanish-English conversation with spontaneous switching. — mix_language.mp3
Output artifact: Output artifact (Image): The bilingual diagnostic shows the engine dropped the Spanish token 'ahora' and rewrote it as 'oh now', with Spanish recall reported at 57.5% (46 of 80 types). — 04-transcript-detail-3.png
What changed: Audio file transformed into Image
Why it matters / Conclusion: It can transcribe mixed-language audio, but Spanish quality is materially weaker than the clean-English medical sample.
The API transcribes mixed-language audio without special configuration, as shown on the Spanish-English sample. The evidence shows successful mixed-language transcription with weaker Spanish recall than clean-English narration.
How it scored on the research's own criteria
The 4 evaluation dimensions from our hands-on research on ElevenLabs Scribe, each judged from recorded runs on 3 test inputs — the same verdicts the ranking page ranks on.
held up partial failed not exercised by this input
| Criterion | Verdict | What the runs showed | Per input | Proof |
|---|---|---|---|---|
| Automation level | Strong5/5 | It runs as one upload request and gives back the transcript payload without any extra operator steps. The integration is as simple as it gets for this kind of task. | open proof ↗ | |
| Export | Strong5/5 | It consistently returns a deep, structured payload rather than a bare transcript. Across all three runs, it included timing, confidence, and speaker information together. | open proof ↗ | |
| Input handling | Strong4/5 | It handled all three uploads and recorded latency, RTF, and cost each time, but one run did not finish cleanly because the API rejected it. That is mostly solid input handling with one meaningful interruption, not a full failure. | open proof ↗ | |
| Output quality | Mixed3/5 | It was excellent on the medical narration, but it lost a lot more ground on overlapping meeting speech and especially on Spanish-English code-switching. The mix of one strong result and two clearly weaker ones lands in the middle. | open proof ↗ |
Verdicts come verbatim from the study's recorded observations, never re-derived at render; a criterion with no recorded run shows Not exercised — this section cannot invent a score.
Developer API pricing
Scribe v2 pricing from the vendor page; the per-hour rate stays flat across paid tiers.
Per-hour Scribe rate does not change across plans: paid tiers buy included hours, not a lower unit rate. Add-ons listed in the report were entity detection (+$0.070/hr) and keyterm prompting (+$0.050/hr).
Banner Preview
How the embed badge will look on your site
Embed HTML
Copy this code to your website source
Quick Integration Guide
- 1Copy the HTML code block above.
- 2Paste it into your site's HTML or CMS editor.
- 3Banner appears instantly on your page.
- 4Links back to your tool profile here.
Similar Tools
Discover more AI tools like ElevenLabs Scribe to enhance your workflow.
Comments (0)
Need a custom AI solution for this use case?
If you are looking to build a custom audio transcription, batch transcription, or transcription workflow for your business or internal workflow, email us at contact@futuresmart.ai.
Found something inaccurate or missing? We try to keep our AI research accurate and useful. If you found outdated information, an issue, or have a suggestion, email us at collaborate@aidemos.com.