A consistent voice per speaker
Profile speakers, retrieve distinct voice references, and convert all turns with SeedVC. The assignment follows the speaker, not the sentence.
All speaker turnsClinical speech · Research frameworkMore than words.
More than words.
More than one
kind of privacy.
Transform voice identity and spoken identifiers separately. Keep the conversation in view. Measure what each intervention changes.
Python framework · Apache-2.0 · Model-adapter architecture
One conversation03 speakers
01 · A
02 · B
03 · A
04 · C
05 · B
06 · A
Two interventions.
One connected system.
Consistent surrogate voices on every turn.
Lexical rendering only where words change.
VOICE IDENTITYIDENTIFIER EDIT
Prosody. Beyond a transcript.Pauses. Part of the interaction.Roles. Who said what.Context. Worth measuring.
The architectureFactorize the change.
Factorize the change.
Follow the evidence.
Changing words and changing a voice solve different problems. The implementation keeps those operations, their artifacts, and their measurements distinct.
Edit the complete affected turn
Plan document-consistent replacements. Render the revised turn with its converted voice reference. Rejoin it with the other converted turns.
Selected complete turnsDo not hide a tradeoff in one score
Inspect speaker linkage, replacement realization, retained identifiers, and clinical preservation at the stage where each was measured.
Separate endpointsThe narrated overview · 2 min 31 sec
Privacy beyond the words.
Why clinical speech needs two privacy interventions, three separate evaluation questions, and governed use.
An illustrated explanation with synthetic narration, not patient footage, model recordings, or a measured before/after demonstration. No autoplay or burned subtitles. Media provenance and rights.
Listen, side by side
Hear the whole conversation.
The original input and the final reconstructed output. Ten complete conversations, from short French dispatch calls to longer English consultations.
Use headphones. Compare the voices, pacing, and delivery. These are public simulated or non-patient conversations, not real-patient recordings. The final output includes both voice conversion and lexical editing.
Full-length playback · 2–15 minutes per recording · Only one player plays at a time
No examples match these filters. Choose another dataset or reset.
01English
PriMock57
Mock primary-care consultation
Original6:33 · full
Final transformed6:32 · full
Recording details 2 detected speakers · 2 edited turns
Source recording: day4_consultation04_mixed. Speaker counts are model-detected, not demographic or reference labels. Final audio uses consistent SeedVC voice conversion, F5-TTS complete-turn lexical editing, and conversation reconstruction.
Original input is a constructed mixture of the doctor and patient role tracks. Dataset source ↗ · Attribution and selection
02English
PriMock57
Mock primary-care consultation
Original9:04 · full
Final transformed9:01 · full
Recording details 2 detected speakers · 1 edited turn
Source recording: day1_consultation03_mixed. Speaker counts are model-detected, not demographic or reference labels. Final audio uses consistent SeedVC voice conversion, F5-TTS complete-turn lexical editing, and conversation reconstruction.
Original input is a constructed mixture of the doctor and patient role tracks. Dataset source ↗ · Attribution and selection
03English
PriMock57
Mock primary-care consultation
Original11:52 · full
Final transformed11:48 · full
Recording details 2 detected speakers · 4 edited turns
Source recording: day3_consultation09_mixed. Speaker counts are model-detected, not demographic or reference labels. Final audio uses consistent SeedVC voice conversion, F5-TTS complete-turn lexical editing, and conversation reconstruction.
Original input is a constructed mixture of the doctor and patient role tracks. Dataset source ↗ · Attribution and selection
04English
Fareez
Simulated respiratory interview
Original8:33 · full
Final transformed8:29 · full
Recording details 2 detected speakers · 6 edited turns
Source recording: RES0217. Speaker counts are model-detected, not demographic or reference labels. Final audio uses consistent SeedVC voice conversion, F5-TTS complete-turn lexical editing, and conversation reconstruction.
Original input is the respiratory interview recording used in the evaluation. Dataset source ↗ · Attribution and selection
05English
Fareez
Simulated respiratory interview
Original11:04 · full
Final transformed10:53 · full
Recording details 2 detected speakers · 1 edited turn
Source recording: RES0068. Speaker counts are model-detected, not demographic or reference labels. Final audio uses consistent SeedVC voice conversion, F5-TTS complete-turn lexical editing, and conversation reconstruction.
Original input is the respiratory interview recording used in the evaluation. Dataset source ↗ · Attribution and selection
06English
Fareez
Simulated respiratory interview
Original12:22 · full
Final transformed12:19 · full
Recording details 2 detected speakers · 2 edited turns
Source recording: RES0123. Speaker counts are model-detected, not demographic or reference labels. Final audio uses consistent SeedVC voice conversion, F5-TTS complete-turn lexical editing, and conversation reconstruction.
Original input is the respiratory interview recording used in the evaluation. Dataset source ↗ · Attribution and selection
07English
Fareez
Simulated respiratory interview
Original15:05 · full
Final transformed14:48 · full
Recording details 3 detected speakers · 2 edited turns
Source recording: RES0011. Speaker counts are model-detected, not demographic or reference labels. Final audio uses consistent SeedVC voice conversion, F5-TTS complete-turn lexical editing, and conversation reconstruction.
Original input is the respiratory interview recording used in the evaluation. Dataset source ↗ · Attribution and selection
08French
SIMSAMU
Simulated emergency-dispatch call
Original2:02 · full
Final transformed2:01 · full
Recording details 2 detected speakers · 1 edited turn
Source recording: dj_2022_avp_scooter. Speaker counts are model-detected, not demographic or reference labels. Final audio uses consistent SeedVC voice conversion, F5-TTS complete-turn lexical editing, and conversation reconstruction.
Original input is the French dispatch-call recording used in the evaluation. Dataset source ↗ · Attribution and selection
09French
SIMSAMU
Simulated emergency-dispatch call
Original2:51 · full
Final transformed2:49 · full
Recording details 4 detected speakers · 1 edited turn
Source recording: dj_2022_douleur_thorax. Speaker counts are model-detected, not demographic or reference labels. Final audio uses consistent SeedVC voice conversion, F5-TTS complete-turn lexical editing, and conversation reconstruction.
Original input is the French dispatch-call recording used in the evaluation. Dataset source ↗ · Attribution and selection
10French
SIMSAMU
Simulated emergency-dispatch call
Original4:17 · full
Final transformed4:17 · full
Recording details 3 detected speakers · 1 edited turn
Source recording: dj_2022_tetanie_2. Speaker counts are model-detected, not demographic or reference labels. Final audio uses consistent SeedVC voice conversion, F5-TTS complete-turn lexical editing, and conversation reconstruction.
Original input is the French dispatch-call recording used in the evaluation. Dataset source ↗ · Attribution and selection
Curated demonstrations, not a performance estimate. Selection spans datasets, languages, lengths, model-detected speaker counts, and edit counts. These examples are not an independent listening study or a guarantee of complete identifier removal.
MP3 files preserve the full source and final timelines. No passages were trimmed for this gallery. Measurements use the underlying WAVs, not these compressed playback files. Selection, provenance, and permissions →
A real starting pointTry the mechanics.
Try the mechanics.
Bring your models.
A lightweight local example verifies reconstruction without a GPU, API key, or model download.
The demo is synthetic tones, not anonymized speech. Real inference requires authorized models, reference voices, and data.
git clone https://github.com/spraphul/aegisvox.git
cd aegisvox
python3 -m venv .venv
source .venv/bin/activate
pip install -e '.[dev,evaluation]'
aegisvox demo --output runs/first-demo
pytest -qFrom inputs to inspectable artifacts.
System architecture ↗Complete-turn contract ↗Models and resident workers ↗Deployment and service API ↗Stage-resolved evaluation ↗Release scope and provenance ↗
A research framework, not a privacy or clinical-safety certificate. Residual linkage, missed identifiers, imperfect edits, and model-based measurement remain important limits.