Evidence · first-party tested/Best AI Tools for Accurate Speech-to-Text on Hard Audio

The tool accepts a 65.39 MB, 2142.709 s audio file and processes it end to end without objection.

✓ Worked🧾 artifact-verifiedinput + output shownTest date not recordedElevenLabs Scribe →

What was measured

Input handling

Whether the tool accepts the benchmark audio as provided and processes it end to end without objection, including the file size/duration it can handle.

context, not decisivecapability

Being able to accept the benchmark audio and handle its size/duration is necessary to test the tool, but it is an operational constraint rather than the transcription quality being ranked. (3 of 3 judges)

What was given, what came back

Test input: Overlapping meeting speech with cross-talk · audio · group: speech-to-text-benchmark

Input — what we sent

Input file 4 — as supplied

0:00 / 0:00

Loading audio...

Overlapping meeting speech with cross-talk

A long AMI meeting audio file with multiple speakers talking over one another, background room noise, and crosstalk. It was used to test how well an STT system handles noisy multi-speaker conversational audio and speaker separation.

Why this input is hard

· overlapping speech
· background noise robustness
· multi-speaker separation
· speaker diarization accuracy
· long-form audio handling

Output — unretouched

Also checked on this input — same tool, 2 other criteria

Export✓ WorkedThe returned payload is rich and fully structured at 3/3 depth, with word-level timestamps, confidence values, speaker labels, and 14,506 timed tokens.Output quality⚠ StruggledTranscript accuracy is weak on crosstalk, with WER 26.67% and 856 substitutions, 782 deletions, and 383 insertions over a 7,579-word reference.

Provenance

Observation: fb95bb1d-6e61-4356-ad55-4b2e77730f17
Evidence run: 469de0c2-d727-4f8f-a60e-e3a5bf8e8588
Study: Transcribe Audio Accurately — Speech-to-Text Engine Benchmark
Research task: 86baxegpu
Tested at: not recorded
Source: first-party
Evidence state: verified
Proof shown: input + output shown
Cost / latency: not captured
Repeat run: not captured
Tester: not captured

The last three rows are honest blanks, not placeholders — our capture has no field for them yet.

Query this

get_evidence({
  tool: "elevenlabs-scribe",
  scenario: "speech-to-text-benchmark"
})

MCP · mcp.aidemos.com/api/mcp

JSON · ai.aidemos.com/v1/observations/fb95bb1d-6e61-4356-ad55-4b2e77730f17

Free with attribution.

Same input, same check — 1 other tool

measured on Input handling

AWS Transcribe✓ WorkedAccepts a 65.39 MB, 2142.709 s WAV and processes it through without objection.

This evidence is published in

Best AI Tools for Accurate Speech-to-Text on Hard Audio →

Real inputs and real outputs, no retouching · every cell queryable via API & MCP · aidemos.com