What it does not do

Summary

Overlapping speech longer than 1.5 s

What happens. When two people talk at once for more than 1.5 s, the words of one of them are dropped or assigned to the other.

Physical cause. One microphone records the sum of both voices, in the same frequency band. A short overlap can be resolved from the clean speech around it. Past 1.5 s there is no clean stretch of either voice to anchor the speaker.

What to do. Record each party on its own channel: stereo call recording, or one microphone per desk in the courtroom. Each channel is then transcribed as its own speaker.

IVR prompts, tones and hold music

What happens. The output gains words nobody said: fragments of the recorded menu, or syllables invented over the music.

Physical cause. Keypad tones and music put steady, harmonic energy in the same 300–3,400 Hz band as the voice. The speech detector cannot tell them apart from speech by level alone, and the recorded IVR voice is speech.

What to do. Trim IVR and hold segments before upload, or export only the agent and customer legs of the call.

Alphanumeric strings without a glossary

What happens. Account numbers, licence plates and case-file numbers come out with letters and digits swapped: “be” for “ve”, “efe” for “ese”.

Physical cause. In Spanish, b and v are the same sound. Letter names are short, and many differ only by a fricative such as /f/ or /s/, whose energy sits mostly above 4 kHz: the upper limit of 8 kHz telephone audio. Without a known format, nothing constrains the string.

What to do. Add the formats to the client glossary: the case-file pattern, the account pattern, the plate pattern. The output is then checked against them.

NOA accent

What happens. On speakers from Salta, Jujuy, Tucumán, Catamarca and Santiago del Estero, WER is higher than the figures on the benchmark.

Physical cause. The NOA assibilated r sounds close to the rioplatense pronunciation of “y” and “ll”, so the model confuses the two. Vowel timing and intonation differ, and the training corpus has few NOA speakers.

What to do. Run the bench test on your own NOA audio before deciding.

Code-switching with Portuguese

What happens. Portuguese words and phrases inside a Spanish conversation are written as the nearest Spanish word.

Physical cause. Portuguese nasal vowels and reduced final vowels have no Spanish category to map to, and the vocabulary is Spanish only.

What to do. Flag those calls for human review. Portuguese is not supported.

Reverberation over RT60 0.8 s

What happens. Errors rise steeply, mostly on word endings and short function words.

Physical cause. RT60 is the time a sound takes to decay by 60 dB. Above 0.8 s, the tail of each syllable is still loud when the next one starts and masks the consonants that separate words. Large rooms with hard walls and floors reach it.

What to do. Put a microphone close to each speaker, such as a lapel or desk microphone, and add absorption to the room. We measure RT60 during the pilot.

Batch only, no streaming

What happens. There are no live captions. A file is transcribed once the recording has ended.

Cause. The model reads each utterance in both directions before it settles a word; it needs the audio that comes after the word. That rules out word-by-word output while someone is still talking.

What to do. At 0.42× real time, each file is ready in less than half its own length. For near-live use, cut recordings into shorter files.