Full-conversation listening examples
Each pair contains the full original input and the full final reconstructed conversation. The final recording follows SeedVC voice conversion, source-language identifier replacement planning, and F5-TTS complete-turn lexical editing. It is not an intermediate voice-conversion output. The unchanged technical film remains silent.
What these examples show
The gallery includes three PriMock57 consultations, four Fareez respiratory interviews, and three SIMSAMU dispatch calls: seven English and three French conversations. Recordings range from roughly two to fifteen minutes. Model-detected speaker counts range from two to four and edited-turn counts from one to six.
Use headphones. Listen to the original and final versions separately, then compare voice identity, delivery, pacing, and conversational continuity. Start each player independently: lexical rendering can change duration, so identical clock times are not guaranteed to refer to identical words. The page pauses other players when a new one starts. Every pair also has direct MP3 and download links.
These are curated demonstrations, not a random evaluation sample, a perceptual quality ranking, or a privacy guarantee. All source recordings are public simulated or non-patient clinical conversations, not recordings from real-patient encounters. Dataset setting and language are confounded and should not be interpreted as a controlled language comparison. No unverified gender, age, or accent labels are shown.
Selection and technical checks
Selection starts with the existing 390-conversation evaluated F5-TTS condition. Eligible recordings have a complete final reconstruction, at least one rendered lexical edit, no failed edited turns, at least two model-detected speakers, a final/pre-edit-converted duration ratio within 0.95 to 1.08, and previously verified sample lineage. There are 204 eligible recordings: 55 PriMock57, 101 Fareez, and 48 SIMSAMU.
We selected examples near the 10th, 50th, and 90th duration percentiles within each eligible dataset, with four Fareez positions near the 10th, 35th, 65th, and 90th percentiles. The middle SIMSAMU position was replaced by a four-detected-speaker example to broaden conversational structure. The exact IDs are in the manifest. No clinical-preservation or replacement-realization score was optimized, and no new human listening score is claimed. This eligibility filter excludes failed reconstructions, so the gallery cannot estimate failure rates.
For every final WAV, we checked its hash against the frozen reconstruction record. All 20 exported files were decoded and checked for positive duration and non-silence. Playback uses 128 kb/s MP3 compression. Gallery export does not trim passages, remove silences, normalize gain, denoise, or alter speed. The input WAV can itself be a processed dataset input, as with the constructed PriMock57 mixtures. Reported study measurements use WAV artifacts, not compressed browser playback.
The public manifest records source IDs, source revisions, original/final durations, file sizes, and MP3 and WAV checksums. These identify the exact examples without disclosing private storage paths or voice references.
Source attribution and permissions
PriMock57
Alex Papadopoulos Korfiatis, Francesco Moramarco, Radmila Sarac, and Aleksandar Savkov;
Babylon Health. PriMock57: A Dataset Of Primary Care Mock Consultations (2022).
Dataset and citation.
Frozen revision: cd2ac707ad03cb4d2531f4ec6b90c659bf4357c5.
CC BY 4.0.
Original gallery inputs are constructed mixtures of the corresponding doctor and patient role tracks. Final versions additionally transform voice identity and edit planned lexical content. Both are compressed to MP3 for playback. These are adaptations; the original dataset creators do not endorse AegisVox. The upstream license is retained in PRIMOCK57-LICENSE.txt.
Fareez respiratory interviews
Faiha Fareez and colleagues, Collection of simulated medical exams. Figshare dataset. CC0 1.0. Source IDs identify recordings in the released respiratory interview collection. Final versions transform voice identity and planned lexical content; browser copies are MP3 exports of the evaluated input and reconstructed WAVs.
SIMSAMU
HeKA Research Team, Simsamu dataset (2023).
Dataset.
Frozen revision: 6b639431f658d636c30dbf342074d723bf303e0f.
MIT License, copyright (c) 2023 HeKA Research Team. The complete copyright and
permission notice is retained in SIMSAMU-LICENSE.txt.
Final versions transform voice identity and planned lexical content; all gallery
files are MP3 playback adaptations.
Transformed-output release scope
The project author confirmed clearance to publish these ten selected transformed outputs on October 12, 2026. This permission is scoped to this gallery; it does not release the reference-voice pool or establish unrestricted rights for other outputs. Upstream dataset rights and notices remain applicable. The repository's Apache-2.0 software license does not relicense these audio recordings. No OpenSwara reference clips, private clinical recordings, reviewer exports, or access codes are included.