Gladia icon
audio-speech

Gladia

Batch STT with rich word-level metadata and strong jargon recall, but overlap handling is weak and bilingual WER needs a mono-downmixed rerun.

Visit Gladia
Batch APIWord-level metadataJargon recall 100%Channel-duplication caveat
TL;DR — our verdictUpdated September 2026 · 9 test artifacts

Strong metadata and jargon performance, but not uniformly reliable

Where it wins
  • You need a batch STT API that returns word-level timestamps, confidence, and speaker labels.
  • You care about technical jargon recall on dense domain audio.
  • You can rerun stereo or code-switching audio with mono downmix or explicit channel control before trusting the bilingual score.
Main limitation
  • You need reliable overlap handling on crosstalk-heavy audio.
Pricing (verified plans)
Starter (pay-as-you-go) Async $0.61/hr · Real-time $0.75/hrGrowth (annual commitment) Async as low as $0.20/hr · Real-time as low as $0.25/hrEnterprise Custom — contact sales
Strongest test artifacts
Batch Transcription →Structured Transcript Output →Robust Speech Transcription →

Our take

Gladia looks strong as a batch STT API for developer workflows: every scored run returned word-level timestamps, confidence, and speaker labels, and the medical-jargon sample scored very well with 100% jargon recall. But the crosstalk case missed large spans, and the bilingual run is confounded by channel duplication, so that WER should not be treated as a clean accuracy result until it is rerun with mono downmix or explicit channel control.

Download video: Demo video
Screen-recording workflow showing the Gladia benchmark setup in Finder, Terminal, and RStudio.

In-Depth Review

Our detailed analysis of Gladia — features, performance, and real-world testing.

AD
AI Demos Team
Expert Reviewer
Verified Review

Feature-by-Feature Breakdown

Batch Transcription
▾
Test Summary
Feature tested: Batch Transcription
Result: Partial

Feature tested: Batch Transcription

Result: Partial

Expected behavior: Completes pre-recorded audio as batch transcription jobs end to end. The capability was exercised on crosstalk.wav, medical_terms.mp3, and mix_language.mp3, with job completion, latency, cost, and transcript outputs observed across those runs.

Test case: Audio file → Image

Input type: Audio file

Input used: Input artifact (Audio file): INPUT — Overlapping Speech / Crosstalk: crosstalk.wav, 65.39 MB, 2142.709 s, mono PCM 16 kHz, diarization enabled. — crosstalk.wav

Observed output: Output artifact (Image): The batch run completed with status scored, 26.33s latency, 0.01229 RTF, and $0.36319 estimated cost; 5,455 words were returned against a 7,579-word reference. — 07-automation-trace-input-1.png

Input artifact: Input artifact (Audio file): INPUT — Overlapping Speech / Crosstalk: crosstalk.wav, 65.39 MB, 2142.709 s, mono PCM 16 kHz, diarization enabled. — crosstalk.wav

Output artifact: Output artifact (Image): The batch run completed with status scored, 26.33s latency, 0.01229 RTF, and $0.36319 estimated cost; 5,455 words were returned against a 7,579-word reference. — 07-automation-trace-input-1.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: medical_terms.mp3, 8.58 MB, 1123.944 s, single-channel narration with diarization enabled. — medical_terms.mp3

Observed output: Output artifact (Image): The batch run completed with status scored, 33.97s latency, 0.03022 RTF, and $0.19051 estimated cost; 2,738 words were returned against a 2,728-word reference. — 07-automation-trace-input-2.png

Input artifact: Input artifact (Audio file): INPUT — Medical Jargon: medical_terms.mp3, 8.58 MB, 1123.944 s, single-channel narration with diarization enabled. — medical_terms.mp3

Output artifact: Output artifact (Image): The batch run completed with status scored, 33.97s latency, 0.03022 RTF, and $0.19051 estimated cost; 2,738 words were returned against a 2,728-word reference. — 07-automation-trace-input-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: mix_language.mp3, 22.19 MB, 1938.495 s, stereo audio with diarization enabled. — mix_language.mp3

Observed output: Output artifact (Image): The batch run completed with status scored, 44.22s latency, 0.02281 RTF, and $0.32857 estimated cost; 10,765 words were returned against a 6,517-word reference. — 07-automation-trace-input-3.png

Input artifact: Input artifact (Audio file): INPUT — Bilingual Code-Switching: mix_language.mp3, 22.19 MB, 1938.495 s, stereo audio with diarization enabled. — mix_language.mp3

Output artifact: Output artifact (Image): The batch run completed with status scored, 44.22s latency, 0.02281 RTF, and $0.32857 estimated cost; 10,765 words were returned against a 6,517-word reference. — 07-automation-trace-input-3.png

What changed: Audio file transformed into Image

Why it matters / Conclusion: The batch workflow itself is solid: all three jobs completed with status scored, and the benchmark captured latency, cost, and returned-word counts for comparison.

Completes pre-recorded audio as batch transcription jobs end to end. The capability was exercised on crosstalk.wav, medical_terms.mp3, and mix_language.mp3, with job completion, latency, cost, and transcript outputs observed across those runs.

audio
0:00 / 0:00
Loading audio...
INPUT — Overlapping Speech / Crosstalk: crosstalk.wav, 65.39 MB, 2142.709 s, mono PCM 16 kHz, diarization enabled.
↓
OUTPUT
Output artifact for "Batch Transcription" test: The batch run completed with status scored, 26.33s latency, 0.01229 RTF, and $0.36319 estimated cost; 5,455 words were returned against a 7,579-word reference., 07-automation-trace-input-1.png
The batch run completed with status scored, 26.33s latency, 0.01229 RTF, and $0.36319 estimated cost; 5,455 words were returned against a 7,579-word reference.
audio
0:00 / 0:00
Loading audio...
INPUT — Medical Jargon: medical_terms.mp3, 8.58 MB, 1123.944 s, single-channel narration with diarization enabled.
↓
OUTPUT
Output artifact for "Batch Transcription" test: The batch run completed with status scored, 33.97s latency, 0.03022 RTF, and $0.19051 estimated cost; 2,738 words were returned against a 2,728-word reference., 07-automation-trace-input-2.png
The batch run completed with status scored, 33.97s latency, 0.03022 RTF, and $0.19051 estimated cost; 2,738 words were returned against a 2,728-word reference.
audio
0:00 / 0:00
Loading audio...
INPUT — Bilingual Code-Switching: mix_language.mp3, 22.19 MB, 1938.495 s, stereo audio with diarization enabled.
↓
OUTPUT
Output artifact for "Batch Transcription" test: The batch run completed with status scored, 44.22s latency, 0.02281 RTF, and $0.32857 estimated cost; 10,765 words were returned against a 6,517-word reference., 07-automation-trace-input-3.png
The batch run completed with status scored, 44.22s latency, 0.02281 RTF, and $0.32857 estimated cost; 10,765 words were returned against a 6,517-word reference.
Bottom Line
The batch workflow itself is solid: all three jobs completed with status scored, and the benchmark captured latency, cost, and returned-word counts for comparison.
From our researchTranscribe Audio Accurately — Speech-to-Text Engine Benchmarkearlier research
Structured Transcript Output
Consistent metadata export across all three inputs.
▾
Test Summary
Feature tested: Structured Transcript Output
Result: Partial — Consistent metadata export across all three inputs.

Feature tested: Structured Transcript Output

Result: Partial

Verdict: Consistent metadata export across all three inputs.

Expected behavior: Returns machine-readable transcript payloads with word-level timing, confidence scores, speaker labels, speaker counts, and related metadata. Across the tested audio, the responses were deep JSON structures that were easy to consume downstream.

Test case: Audio file → Image

Input type: Audio file

Input used: Input artifact (Audio file): INPUT — Overlapping Speech / Crosstalk: diarization enabled on AMI EN2002a. — crosstalk.wav

Observed output: Output artifact (Image): The raw response reported speaker_labels yes and distinct Speakers 4, matching the four participants in the crosstalk dataset. — 02-response-raw-input-1.png

Input artifact: Input artifact (Audio file): INPUT — Overlapping Speech / Crosstalk: diarization enabled on AMI EN2002a. — crosstalk.wav

Output artifact: Output artifact (Image): The raw response reported speaker_labels yes and distinct Speakers 4, matching the four participants in the crosstalk dataset. — 02-response-raw-input-1.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: diarization enabled on single-speaker narration. — medical_terms.mp3

Observed output: Output artifact (Image): The raw response reported speaker_labels yes and distinct Speakers 1 on the medical narration sample. — 02-response-raw-input-2.png

Input artifact: Input artifact (Audio file): INPUT — Medical Jargon: diarization enabled on single-speaker narration. — medical_terms.mp3

Output artifact: Output artifact (Image): The raw response reported speaker_labels yes and distinct Speakers 1 on the medical narration sample. — 02-response-raw-input-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: diarization enabled on stereo audio. — mix_language.mp3

Observed output: Output artifact (Image): The raw response reported speaker_labels yes, distinct speakers 5, and distinct channels 2 on the bilingual stereo sample. — 02-response-raw-input-3.png

Input artifact: Input artifact (Audio file): INPUT — Bilingual Code-Switching: diarization enabled on stereo audio. — mix_language.mp3

Output artifact: Output artifact (Image): The raw response reported speaker_labels yes, distinct speakers 5, and distinct channels 2 on the bilingual stereo sample. — 02-response-raw-input-3.png

What changed: Audio file transformed into Image

Why it matters / Conclusion: The output shape is stable and integration-friendly: word timestamps, confidence, and speaker labels are always present in the scored runs.

Returns machine-readable transcript payloads with word-level timing, confidence scores, speaker labels, speaker counts, and related metadata. Across the tested audio, the responses were deep JSON structures that were easy to consume downstream.

audio
0:00 / 0:00
Loading audio...
INPUT — Overlapping Speech / Crosstalk: diarization enabled on AMI EN2002a.
↓
OUTPUT
Output artifact for "Structured Transcript Output" test: The raw response reported speaker_labels yes and distinct Speakers 4, matching the four participants in the crosstalk dataset., 02-response-raw-input-1.png
The raw response reported speaker_labels yes and distinct Speakers 4, matching the four participants in the crosstalk dataset.
audio
0:00 / 0:00
Loading audio...
INPUT — Medical Jargon: diarization enabled on single-speaker narration.
↓
OUTPUT
Output artifact for "Structured Transcript Output" test: The raw response reported speaker_labels yes and distinct Speakers 1 on the medical narration sample., 02-response-raw-input-2.png
The raw response reported speaker_labels yes and distinct Speakers 1 on the medical narration sample.
audio
0:00 / 0:00
Loading audio...
INPUT — Bilingual Code-Switching: diarization enabled on stereo audio.
↓
OUTPUT
Output artifact for "Structured Transcript Output" test: The raw response reported speaker_labels yes, distinct speakers 5, and distinct channels 2 on the bilingual stereo sample., 02-response-raw-input-3.png
The raw response reported speaker_labels yes, distinct speakers 5, and distinct channels 2 on the bilingual stereo sample.
Bottom Line
The output shape is stable and integration-friendly: word timestamps, confidence, and speaker labels are always present in the scored runs.
From our researchTranscribe Audio Accurately — Speech-to-Text Engine Benchmarkearlier research
Robust Speech Transcription
Weak on crosstalk-heavy audio.
▾
Test Summary
Feature tested: Robust Speech Transcription
Result: Failed — Weak on crosstalk-heavy audio.

Feature tested: Robust Speech Transcription

Result: Failed

Verdict: Weak on crosstalk-heavy audio.

Expected behavior: Handles difficult speech conditions such as overlapping speakers, domain-specific jargon, and Spanish-English code-switching. The capability was exercised on crosstalk-heavy meeting audio, a medical lecture, bilingual files, and speaker-labeled overlap variants.

Test case: Audio file → Image

Input type: Audio file

Input used: Input artifact (Audio file): INPUT — Overlapping Speech / Crosstalk: four-way meeting audio from AMI EN2002a, 65.39 MB, 2142.709 s. — crosstalk.wav

Observed output: Output artifact (Image): The transcript detail shows 37.35% WER with 529 substitutions, 2213 deletions, and 89 insertions against a 7,579-word reference; the engine detected 4 speaker labels, matching the four participants, but the transcript is incomplete. — 04-transcript-detail-input-1.png

Input artifact: Input artifact (Audio file): INPUT — Overlapping Speech / Crosstalk: four-way meeting audio from AMI EN2002a, 65.39 MB, 2142.709 s. — crosstalk.wav

Output artifact: Output artifact (Image): The transcript detail shows 37.35% WER with 529 substitutions, 2213 deletions, and 89 insertions against a 7,579-word reference; the engine detected 4 speaker labels, matching the four participants, but the transcript is incomplete. — 04-transcript-detail-input-1.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: Gray's Anatomy narration, 8.58 MB, 1123.944 s, single speaker with medical terminology. — medical_terms.mp3

Observed output: Output artifact (Image): The transcript detail shows 4.07% WER with 73 substitutions, 14 deletions, and 24 insertions against a 2,728-word reference, and jargon recall was 100.0%. — 04-transcript-detail-input-2.png

Input artifact: Input artifact (Audio file): INPUT — Medical Jargon: Gray's Anatomy narration, 8.58 MB, 1123.944 s, single speaker with medical terminology. — medical_terms.mp3

Output artifact: Output artifact (Image): The transcript detail shows 4.07% WER with 73 substitutions, 14 deletions, and 24 insertions against a 2,728-word reference, and jargon recall was 100.0%. — 04-transcript-detail-input-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: Bangor Miami spontaneous Spanish-English conversation, 22.19 MB, 1938.495 s, stereo audio. — mix_language.mp3

Observed output: Output artifact (Image): The transcript detail shows 88.45% WER, Spanish recall 56.2%, and an explicit multi-channel duplication caveat: the response carried 2 distinct channel values and 10,765 transcript words against a 6,517-word reference, so the WER is an artefact rather than a clean accuracy result. — 04-transcript-detail-input-3.png

Input artifact: Input artifact (Audio file): INPUT — Bilingual Code-Switching: Bangor Miami spontaneous Spanish-English conversation, 22.19 MB, 1938.495 s, stereo audio. — mix_language.mp3

Output artifact: Output artifact (Image): The transcript detail shows 88.45% WER, Spanish recall 56.2%, and an explicit multi-channel duplication caveat: the response carried 2 distinct channel values and 10,765 transcript words against a 6,517-word reference, so the WER is an artefact rather than a clean accuracy result. — 04-transcript-detail-input-3.png

What changed: Audio file transformed into Image

Why it matters / Conclusion: It recognized the number of speakers, but overlap-heavy speech still produced a badly incomplete transcript.

Handles difficult speech conditions such as overlapping speakers, domain-specific jargon, and Spanish-English code-switching. The capability was exercised on crosstalk-heavy meeting audio, a medical lecture, bilingual files, and speaker-labeled overlap variants.

audio
0:00 / 0:00
Loading audio...
INPUT — Overlapping Speech / Crosstalk: four-way meeting audio from AMI EN2002a, 65.39 MB, 2142.709 s.
↓
OUTPUT
Output artifact for "Robust Speech Transcription" test: The transcript detail shows 37.35% WER with 529 substitutions, 2213 deletions, and 89 insertions against a 7,579-word reference; the engine detected 4 speaker labels, matching the four participants, but the transcript is incomplete., 04-transcript-detail-input-1.png
The transcript detail shows 37.35% WER with 529 substitutions, 2213 deletions, and 89 insertions against a 7,579-word reference; the engine detected 4 speaker labels, matching the four participants, but the transcript is incomplete.
audio
0:00 / 0:00
Loading audio...
INPUT — Medical Jargon: Gray's Anatomy narration, 8.58 MB, 1123.944 s, single speaker with medical terminology.
↓
OUTPUT
Output artifact for "Robust Speech Transcription" test: The transcript detail shows 4.07% WER with 73 substitutions, 14 deletions, and 24 insertions against a 2,728-word reference, and jargon recall was 100.0%., 04-transcript-detail-input-2.png
The transcript detail shows 4.07% WER with 73 substitutions, 14 deletions, and 24 insertions against a 2,728-word reference, and jargon recall was 100.0%.
audio
0:00 / 0:00
Loading audio...
INPUT — Bilingual Code-Switching: Bangor Miami spontaneous Spanish-English conversation, 22.19 MB, 1938.495 s, stereo audio.
↓
OUTPUT
Output artifact for "Robust Speech Transcription" test: The transcript detail shows 88.45% WER, Spanish recall 56.2%, and an explicit multi-channel duplication caveat: the response carried 2 distinct channel values and 10,765 transcript words against a 6,517-word reference, so the WER is an artefact rather than a clean accuracy result., 04-transcript-detail-input-3.png
The transcript detail shows 88.45% WER, Spanish recall 56.2%, and an explicit multi-channel duplication caveat: the response carried 2 distinct channel values and 10,765 transcript words against a 6,517-word reference, so the WER is an artefact rather than a clean accuracy result.
Bottom Line
It recognized the number of speakers, but overlap-heavy speech still produced a badly incomplete transcript.
From our researchTranscribe Audio Accurately — Speech-to-Text Engine Benchmarkearlier research

How it scored on the research's own criteria

The 3 evaluation dimensions from our hands-on research on Gladia, 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

CriterionVerdictWhat the runs showedPer inputProof
Output qualityMixed3/5It is excellent on the medical narration, but the other two cases are weak for different reasons: one drops a lot of overlapping speech, and the bilingual case inflates the transcript by repeating both channels. That split makes the overall transcription quality mixed rather than reliably strong.open proof ↗
Automation levelStrong4/5The runs themselves finish on their own, but the workflow still needs a human to pick the engine and provide a key before that happens. So it is mostly automated, not fully hands-off from start to finish.open proof ↗
Input handlingStrong5/5It took all three files without objection and finished them well faster than real time, with costs staying in the expected list-price range. That consistency across very different audio types is what earns the top score.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.

Official pricing

Async and real-time rates differ, and the paid plans bundle the core speech features used in this benchmark.

Starter (pay-as-you-go)
Async $0.61/hr · Real-time $0.75/hr
€50 free credits one-time; 25 async concurrent requests; 30 real-time concurrent; no uptime SLA or priority queue.
Growth (annual commitment)
Async as low as $0.20/hr · Real-time as low as $0.25/hr
Flexible concurrency; custom volume discounts; 99.9% uptime SLA; priority queue; model-training opt-out.
Enterprise
Custom — contact sales
Unlimited concurrency; zero data retention; SLAs; custom hosting; custom models and fine-tuning.

Source: https://www.gladia.io/pricing, accessed 2026-08-14. The report notes that paid plans include diarization, automatic language detection/switching, word-level timestamps, and 100+ languages; Enterprise adds zero data retention and custom hosting.

✓ Use This If
●You need a batch STT API that returns word-level timestamps, confidence, and speaker labels.
●You care about technical jargon recall on dense domain audio.
●You can rerun stereo or code-switching audio with mono downmix or explicit channel control before trusting the bilingual score.
✕ Skip This If
●You need reliable overlap handling on crosstalk-heavy audio.
●You need a trustworthy bilingual WER without first rerunning channel-controlled audio.
●You need one of the cheapest options; this benchmark ranked Gladia 9/10 on price.
audio-speechaudio-to-textspeechOther
Yes. Every scored run showed a stable developer payload with word_timestamps, confidence, and speaker_labels present, and the raw responses reported payload depth 3/3.
Very good. On the Gray's Anatomy sample it scored 4.07% WER, returned 2,738 words against a 2,728-word reference, and achieved 100.0% recall on the scored jargon terms.
Poorly on transcript completeness. On the crosstalk-heavy AMI clip it scored 37.35% WER with 2,213 deletions, though it did detect 4 speakers, matching the true participant count.
No, not as recorded here. The bilingual run carried 2 distinct channel values and doubled the transcript length, so the 88.45% WER is an artefact of channel duplication. It should be rerun with mono downmix or explicit channel control before being cited.
The vendor pricing page lists Starter at async $0.61/hr and real-time $0.75/hr, Growth as low as async $0.20/hr and real-time $0.25/hr, and Enterprise as custom pricing.
No. This was a batch benchmark, and the report says per-call timings for the multi-stage HTTP flow were not instrumented when the run executed.

Banner Preview

How the embed badge will look on your site

Gladia featured on AI Demos

Embed HTML

Copy this code to your website source

<a target="_blank" href="https://aidemos.com/tools/gladia?utm_source=gladia_embed" style="width: 250px; height: 80px; border-radius:4px;" width="250" height="80"> <img src="https://aidemos-website-images.s3.amazonaws.com/featured.png" alt="Gladia | Featured on AI Demos" style="width: 250px; height: 80px; border-radius:4px;" width="250" height="80"> </a>

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

Similar Tools

Discover more AI tools like Gladia to enhance your workflow.

🤖
Deepgram Speech-to-Text
AI Tool
🤖
AssemblyAI
Fast batch speech-to-text with rich metadata, strong jargon and mixed-language results, but overlap-heavy meetings can still lose too much.
AI Tool
🤖
Speechmatics
Strong batch STT for hard English audio, but weak on code-switching as configured.
AI Tool
🤖
OpenAI Whisper
AI Tool
🤖
Google Cloud Speech-to-Text
Timed batch transcripts for mostly English, jargon-heavy audio — but not for diarization or code-switching.
AI Tool
🤖
AWS Transcribe
Batch speech-to-text with word-level metadata, but accuracy drops on overlap and code-switching.
AI Tool
🤖
ElevenLabs Scribe
Fast batch speech-to-text with word-level metadata, strongest on jargon and weaker on overlap/code-switching.
AI Tool
🤖
Rev AI
Low-cost batch speech-to-text with structured JSON, word timestamps, and speaker labels, but mixed accuracy on crosstalk and code-switching.
AI Tool
🤖
GroqCloud Speech-to-Text
AI Tool
🤖
Deepgram
Batch speech-to-text with word-level metadata and speaker labels, but weak on crosstalk and code-switching.
AI Tool
🤖
OpenAI Whisper-1
AI Tool
🤖
Google Cloud STT v2
AI Tool
🤖
GroqCloud Whisper Large-v3
AI Tool
🤖
OpenAI
Batch speech-to-text with word timestamps, but a strict upload cap and weak multilingual performance make it a mixed fit for hard audio.
AI Tool
🤖
ElevenLabs
Natural-sounding voice cloning and narration, but with only approximate voice identity.
AI Tool
🤖
Deepgram Nova-3
AI Tool
🤖
AssemblyAI Universal
AI Tool
🤖
Speechmatics Ursa Enhanced
AI Tool

Comments (0)

Please Log in to join the discussion.

Built by FutureSmart AI — the team behind AI Demos

Need a custom AI solution for this use case?

If you are looking to build a custom speech-to-text transcription, audio transcription, or batch transcription pipeline for your business or internal workflow, email us at contact@futuresmart.ai.

Get a custom build

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.

Back to Top