Voice resemblance improved before call readiness did
Seed-VC2 converted existing Supertonic speech with four, eight and twelve diffusion steps. The selected public audiobook references were thirty seconds long, but the upstream wrapper used only the first twenty-five seconds.
Key finding
Timbre-only transfer improved speaker-embedding similarity on short greetings, while style transfer could destroy the words.
Edited finding from the pinned experiment record. The report and figure metadata preserve the original workload and limits.
The boundary
Public audiobook demo references, tiny greeting corpus and offline conversion. Similarity is not identity probability or human naturalness.
Write-ups
Report tts/seed-vc2-voice-transfer-2026-10-13/README.md
A voice-conversion stage after synthesis
Seed-VC2 converted existing Supertonic speech with four, eight and twelve diffusion steps. The selected public audiobook references were thirty seconds long, but the upstream wrapper used only the first twenty-five seconds.
Do not optimize one score
A short English style/accent conversion reached speaker cosine around 0.880, yet local ASR recorded 100% word error on its six-word greeting. Timbre-only conversion preserved the greeting while improving resemblance. A successful conversion request therefore did not mean usable translated speech.
A separate female-reference test completed twelve conversions. Four-step timbre-only English measured cosine 0.861, about 1.119 seconds to the first complete chunk and zero ASR edits on six words. Spanish measured 0.803, 1.627 seconds and zero edits on five words. These tiny samples cannot establish production intelligibility.
Keep it experimental
The source study used demo narrators; no family voice is published here. Adding conversion to measured translation-stage timings yields only an arithmetic latency estimate. Real two-phone calls, consented enrollment, interruptions and network delay were not validated by these offline tests.
Evidence and publication boundary
This is an edited account of the recorded experiment. The linked archive is private; the methods, aggregate findings and limits are included above. Reference recordings, raw conversations, authentication material and workstation configuration are excluded.
More in Speech & codecs
MOSS on a 3090: Q2 failed, AWQ4 stayed closer
Every profile began with the pinned BF16 generator. The comparison reused fixed reference codes and an FP32 decoder for the main quality measurements, keeping codec changes separate from generator quantization.
AWQ4 stayed close to the BF16 reference on WavLM and character errors; the sensitivity Q2 candidate was rejected.
Drama3: BF16 decoded the same codes about three times faster
The experiment replayed saved Spanish and English codes representing 21.5481 and 15.8824 seconds of audio. It compared FP32, BF16, mixed BF16 and two INT8 decoder variants on one RTX 3090.
BF16 reduced median decoder time from 390.07 to 129.54 ms in Spanish and from 292.42 to 96.88 ms in English.
Keeping FP32 and moving Supertonic to Triton
Rather than reduce the CPU graph precision again, this follow-up used NVIDIA Triton’s ONNX backend with CUDA execution and the unchanged FP32 graphs. The paired comparison retained both speed and acoustic checks.
The paired GPU test cut median first-audio time to about 0.35 seconds while preserving the measured ASR results.