Prosodia

Product

Prosodia Local

Speech recognition for rioplatense Spanish that runs on your server. Batch transcription with speakers, timestamps and your glossary. Nothing leaves your network.

Prosodia Local transcription console showing a hearing transcript. Each turn has a start and end timestamp, and each of three speakers has its own color and label: judge, witness and defense.
Transcription console: one row per turn, speaker and timestamps on the left.

Minimum hardware

GPUOne NVIDIA L4 (24 GB) or better per server
Speed0.42× real time on one L4: one hour of audio in about 25 minutes
CPU, memory, diskSized during the pilot from your daily audio volume. We send the bill of materials before installation.
Operating systemLinux x86-64 (Ubuntu LTS or Red Hat Enterprise Linux) with the NVIDIA driver
NetworkNo outbound connection. Local network only, for the API and the shared folder.

Audio formats

ContainersWAV (PCM), FLAC, MP3, OGG/Opus
Sample rate8 kHz (telephone) and 16 kHz (room) natively. Other rates are resampled to 16 kHz.
ChannelsMono or stereo. In stereo, each channel can be transcribed as its own speaker.
ModeBatch only. No live streaming; see limits.

Glossary per client

Each client gets its own glossary, loaded before the first file. It holds three kinds of entries:

  • Terms: product names, surnames, street names, legal vocabulary.
  • Spoken variants: how each term is actually said, including abbreviations and nicknames.
  • Formats: patterns for case-file numbers, account numbers and plates. The output is checked against them.

Glossaries are versioned. Each transcript records the glossary version it used, so a result can be reproduced.

Glossary view: a table of client terms with columns for the written form, its spoken variants and the entry type, plus a format pattern for case-file numbers.
Glossary view: written form, spoken variants and type for each entry.

Diarization

Diarization answers who spoke when. Prosodia Local separates up to 6 speakers per recording and labels each turn with its start and end time.

Speakers in the recordingDER
28.7%
Up to 614.2%

Speakers come out as S1 to S6. You can rename them (agent, customer, judge, witness) and the names carry into every export. Definitions and protocol are on the benchmark.

Export

FormatContents
TXTPlain text, one turn per line with speaker and start time
DOCXDocument with speakers and timestamps, ready to edit
SRT, WebVTTSubtitles for playback next to the audio
JSONWords with start and end times, confidence and speaker, for your own systems

Integration: folder or API

Folder

Drop audio files into a watched folder on the server or on a network share. Transcripts appear in the output folder, in the formats you chose. No code needed.

API

HTTPS on your local network, with one token per client application. Send a file, then read the job until it is done.

# Example request. Host and token are set at installation.
curl -H "Authorization: Bearer $TOKEN" \
     -F "audio=@hearing.wav" -F "glossary=juzgado-civil-3" \
     https://prosodia.internal/api/v1/jobs

curl -H "Authorization: Bearer $TOKEN" \
     https://prosodia.internal/api/v1/jobs/$JOB_ID

Server requirements, hardening and offline updates are on Pricing and deployment.

Module

Acta, for courts

Acta turns the recording of a hearing into a draft record, in the court’s own format. It is in pilot at one court.

  • Header fields filled from the case: case file, case title, court and date.
  • Speakers labelled by procedural role: judge, prosecutor, defense, witness.
  • A timestamp on every turn, so each passage can be checked against the audio.
  • A SHA-256 hash of the source audio printed in the record, to show the file was not altered.
  • Export to DOCX on the court’s template.

Acta produces a draft. The clerk reviews it and signs it; the system does not sign anything.