How much audio was needed to make the clone?

The reference was a single 15.3-second recording of the person reading a passage different from the test script.

Did the clone handle emotional instructions like 'panicking'?

The clone's speaking speed reached 8.40 versus the real person's 8.54 under the panicked instruction, but its pitch stayed around 93Hz instead of rising.

What tools measured the clone's similarity?

Similarity was calculated with WavLM-large and ECAPA-TDNN speaker embeddings using cosine similarity, with models from k2-fsa/TTS_eval_models.

Audio & SpeechQiita 機械学習1d agoNov 4, 2026, 10:00 JST

VoiceStudio clone scores 0.809 with VoxCPM2 in voice test

Send on LINE

3 Key Points

What happened
Using VoiceStudio with the VoxCPM2 model, a 15.3-second sample produced a clone whose normal-speech similarity to the real person was 0.809; the app's OmniVoice model scored 0.825.
Why it matters
On ordinary speaking styles, such short-sample clones can already land near the real speaker's recordings, though the test used only one person and three runs per condition.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, every morning.

Get these free every morning

Context & Analysis

The test was run entirely on a Mac with an Apple M4 chip and 16GB of memory, using VoiceStudio v0.5.7 launched from source, with three generations per condition at seeds 0, 1, and 2. VoxCPM2 is described as able to follow speaking-style instructions in words, while OmniVoice does not support emotional instructions and was used only for the normal-speaking comparison.

The similarity metric is not a pure measure of voice quality. When the same person read in different emotional styles, similarity to the reference recording fell from 0.881 for normal speech to 0.629 for panicked and 0.600 for timid. That suggests the roughly 0.50 scores for instructed clones partly reflect a different reading style rather than the voice becoming a different person.

The author also notes that for VoxCPM2 the no-instruction and with-instruction clones used different generation methods in VoiceStudio, so comparisons between them mix the effect of the instruction with that method difference. License terms differ too: VoxCPM2 is listed as Apache 2.0 on its model card, while OmniVoice's code is Apache 2.0 but its trained weights are CC-BY-NC (non-commercial).

FAQ

How much audio was needed to make the clone?: The reference was a single 15.3-second recording of the person reading a passage different from the test script.
Did the clone handle emotional instructions like 'panicking'?: The clone's speaking speed reached 8.40 versus the real person's 8.54 under the panicked instruction, but its pitch stayed around 93Hz instead of rising.
What tools measured the clone's similarity?: Similarity was calculated with WavLM-large and ECAPA-TDNN speaker embeddings using cosine similarity, with models from k2-fsa/TTS_eval_models.

Qiita 機械学習Read Original Article

Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Get it free

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.