Transcribe rioplatense Spanish on your own server.
Prosodia Local turns calls and court hearings into transcripts with speakers and timestamps.
It runs inside your network: audio and text never leave it.
- 11.4%WER on 8 kHz telephone audio
- 0.42×real time on one NVIDIA L4
- Up to 6speakers per recording, by diarization
Audio samples
Three 5-second clips. The human reference and the model output, side by side.
Unedited audio, as the model receives it, from our own recordings with each speaker’s consent; no customer audio. Every error is marked and counted.
-
Reference
Mirá, yo ya llamé tres veces esta semana y nadie me resolvió lo de (deleted by the model) la tarjeta.
Prosodia Local
Mirá, yo ya llamé tres veces esta semana y nadie me resolvió two words missing la tarjeta.
2 deletions in 16 reference words. “lo de” was spoken as a single syllable.
-
Reference
Que conste en el acta que el testigo reconoce la firma obrante a fojas (substituted by the model) doce.
Prosodia Local
Que conste en el acta que el testigo reconoce la firma obrante a hojas (substitution for “fojas”) doce.
1 substitution in 15 reference words. The /f/ of “fojas” is a weak sound and was masked by room noise.
-
Reference
No, no, no, el titular es mi viejo, nothing here in the reference yo soy el hijo, ¿me entendés?
Prosodia Local
No, no, no, el titular es mi viejo y (inserted by the model) yo soy el hijo, ¿me entendés?
1 insertion in 14 reference words. “Yo” was pronounced [ʃo]; the model added a “y” before it. The voseo “entendés” is transcribed as spoken.
- S substitution
- D deletion
- I insertion
- Case and punctuation are not scored.
The problem
Why general models fail on this audio
Neutral models miss voseo and elision.
“Tenés”, “vení”, “pa’ que”, an aspirated s. These forms are rare in neutral training data, so the errors land on the words that carry the meaning.
Court audio cannot go to a foreign cloud.
Hearings record parties and witnesses. Ley 25.326 limits sending personal data to other countries. Transcription has to run where the audio is stored.
The client’s glossary is not optional.
Product names, case-file formats, street names, the judges’ surnames. Without them a model writes the nearest common word. We load the glossary before the first file.
How it works
Four steps, all inside your network
- We install Prosodia Local on a server in your room.
- We load your glossary: names, products, case-file and account formats.
- You transcribe by dropping files in a folder or by calling the API.
- You export text, subtitles or JSON, with speakers and timestamps.
Performance
Word error rate by recording type
WER is the share of words wrong against a human reference; lower is better. Both systems get the same audio and the same scoring.
| Recording type | Prosodia Local | Whisper large-v3 |
|---|---|---|
| Telephone, 8 kHz | 11.4% | 28.3% |
| Courtroom, 16 kHz | 6.2% | Not measured yet |
| Clean read speech | 4.1% | 5.1% |
What a customer says
We used to audit a sample of the calls. Now quality searches every call, and none of the audio leaves our building.
Bench test
Send 20 minutes of your own audio. Get your error rate back.
- Pick 20 minutes that represent your calls or hearings, with a human transcript of the same audio.
- We return our transcript, the WER against your transcript and every error, listed.
- If the audio cannot leave your building, we run the test there.