Datasheet · 2026-11
Prosodia Local
On-premise speech recognition for rioplatense Spanish. Batch transcription with diarization and a client glossary.
Print this page or save it as PDF from your browser; it fits on one A4 sheet.
Word error rate
| Recording type | Prosodia | Whisper large-v3 |
|---|---|---|
| Telephone, 8 kHz | 11.4% | 28.3% |
| Courtroom, 16 kHz | 6.2% | Not measured |
| Clean read speech | 4.1% | 5.1% |
Evaluation set: 15 h (6 h Common Voice es-AR, 9 h own), split by speaker. Corpus: 1,900 h (214 h own, 1,686 h licensed).
Diarization and speed
| DER, 2 speakers | 8.7% |
|---|---|
| DER, up to 6 speakers | 14.2% |
| Maximum speakers | 6 |
| Speed on one NVIDIA L4 | 0.42× real time |
Hardware and system
| GPU | One NVIDIA L4 (24 GB) or better |
|---|---|
| CPU, memory, disk | Sized during the pilot |
| OS | Linux x86-64, Ubuntu LTS or RHEL |
| Network | No outbound connection |
| Updates | Signed offline packages, with rollback |
| Compliance | SOC 2 Type II audit in progress |
Input, output, integration
| Audio | WAV, FLAC, MP3, OGG/Opus; mono or stereo |
|---|---|
| Sample rate | 8 kHz and 16 kHz native; others resampled |
| Export | TXT, DOCX, SRT, WebVTT, JSON |
| Integration | Watched folder or HTTPS API on the local network |
| Glossary | Per client: terms, spoken variants, formats |
| Courts | Acta module, in pilot |
Limits
- Overlapping speech longer than 1.5 s: words dropped or assigned to the wrong speaker.
- IVR prompts, keypad tones and hold music: spurious words.
- Alphanumeric strings without a glossary: letters and digits substituted.
- NOA accent: higher WER than the figures above.
- Code-switching with Portuguese: not supported.
- Reverberation over RT60 0.8 s: word endings and short words lost.
- Batch only: no live streaming.
Causes and workarounds: prosodia.com.ar/limits/
Contact
Prosodia Sistemas SAS · CUIT 30-71890412-5 · Av. Wheelwright 1200, Piso 4, S2000 Rosario, Santa Fe, Argentina · hola@prosodia.com.ar · +54 341 555-6180