Ground the comparison in audio.
Compare what was transcribed with what was said in this recording.
BESSEL · SPEECH DECISION API
Rank transcripts.
Against the audio.
Send a recording and its candidate transcripts. Bessel compares them against the audio and returns a ranking your application can use.
AUDIO + TRANSCRIPTS → RANKINGRecorded example
speech-sample.mp39s
0:000:030:060:09
CANDIDATE / SOURCERELATIVE SCORE
01
Azure Speech Fast
0.82902
Deepgram Nova-3
0.171Azure Speech Fast
“…that is not the story that the remembering self keeps for us.”
Deepgram Nova-3
“…that is not the story that our remembering self keeps for us.”
Bring your own candidates.
Use transcripts from any ASR provider, your own model or multiple decoding attempts.
Keep your code in control.
Use candidate IDs, relative scores and explicit uncertainty to decide your next step.
For teams building
with speech.
FOR PRODUCT TEAMS
Choose a transcript.
Add an audio-based comparison to the speech pipeline you already run.
FOR MODEL TEAMS
Improve your next model.
Compare candidate outputs from model versions or decoding strategies on the same clip.
Try Bessel.
AWAITING INPUT
Your results
PROCESSING PROGRESS0%
- Upload
- Queue
- Transcribe
- Rank
Live updates0:00 elapsed
MODEL RANKINGSCORE
About this result
Scores compare the selected models; they are not accuracy percentages. This comparison does not change your ASR.
API guide ↗View result JSON
No comparison yet.
Questions, answered.
What does Bessel return?
The ranking API returns your candidate IDs in rank order, relative scores and a recommendation only when the comparison has enough support. It does not generate a new transcript.
Can Bessel lower my word error rate?
Choosing a stronger transcript can reduce WER. A ranking is relative to your candidates: even the highest-ranked transcript can contain errors. Measuring WER requires a reference transcript.
Which transcripts can I compare?
Bring 2–4 transcripts of the same 0.1–15 second clip, from any source. Ranking supports 1,650 languages. In the playground, upload audio and choose from 29 ASR models; Bessel generates the transcripts for you. Recordings up to 5 minutes are supported there; transcription languages depend on the selected ASR models. See language coverage.
Do I need to train Bessel on my audio?
No. Submit a recording and its candidate transcripts. You do not need to train Bessel on your audio first.