A small voice model ran on both machines. Training revealed the real risk.

Audio8 fit on a 16 GB Mac and an RTX 3090 Ti. Its lowest-loss checkpoint still failed to stop speaking.

By Wei Jie Chee · 30 Jul 2026

Read the printable research note
16 GBMac inference worked
2.9 GBhighest observed GPU use
100 stepsthe checkpoint we kept
Named-model investigator + Voice-quality evaluatorCompare checkpoints and voices by consequential failures.

Can a small cross-platform TTS model learn a usable voice?

Audio8 ran on both tested machines, but later lower-loss checkpoints often failed to stop speaking.

Do this first

Judge end-of-speech behavior and held-out audio before selecting the checkpoint with the lowest logged loss.

Keep this boundary in view

One model revision, corpus slice, Mac, GPU, and checkpoint sweep do not establish general Audio8 quality or stability.

Inspect the evidence and what would change the answer

First-party cross-platform run and checkpoint sweep

What supports it

The experiment records runtime, memory, training, ASR, speaker-similarity proxies, and end-of-speech behavior across checkpoints.

Change our mind if

A corrected or repeated run makes later checkpoints end reliably while improving held-out listening quality.

Claim TTS-AUDIO8-END-2026-08 / v1, checked 7 Aug 2026

The loss kept falling. The voice stopped ending.

Step 500 looked strongest in the training log. Generation exposed the opposite: later checkpoints often continued until the 500-frame safety cap. A usable voice must pass speech behavior, not only validation loss.

  1. 100selected
  2. 150lower loss
  3. 200lower loss
  4. 250lower loss
  5. 300lower loss
  6. 350lower loss
  7. 400lower loss
  8. 450lower loss
  9. 500lowest loss
ended normally hit the safety cap

The next result must be heard, not inferred.

ASR recovered the short test sentences and a speaker-embedding proxy moved toward the target profile. Neither measurement establishes naturalness or listener preference. The private comparison remains a review gate, not public evidence on this page.

60-second takeaway

  • It ran on the 16 GB M2 Mac. CPU and MPS both produced valid 44.1 kHz audio without swap activity in our short tests.
  • It fit easily on the RTX 3090 Ti. The highest observed whole-GPU memory reading was 2,938 MB during a two-item English and Chinese batch.
  • The short speech was textually clear. Whisper large-v3-turbo recovered all seven tested outputs with 0.0 WER or CER after punctuation-insensitive normalization.
  • MPS was not a clear speed win. It finished slightly sooner end to end, but its measured generation loop was not faster than CPU in these single short runs.
  • Voice cloning worked, but the quality verdict remains open. We used a synthetic macOS reference rather than a person's voice. ECAPA speaker similarity landed between 0.564 and 0.604, which is a diagnostic signal rather than a human listening score.
  • The 128-clip LoRA pilot moved toward the target voice. Mean ECAPA cosine similarity against 16 paired FEMALE_01 recordings increased from 0.092 for the base model to 0.189 after 100 steps.
  • Lower validation loss hid a serious failure. A 500-step run kept improving on validation loss, but later checkpoints often failed to stop speaking. Step 300 failed to end normally on 12 of 16 prompts.
  • The model is promising, not production-proven. Blind listening, larger clean datasets, long passages, and broader checkpoint tests remain necessary.

What Audio8 TTS Preview 0.6B is

Audio8 uses a DualAR design inspired by Fish Audio S2 Pro.

The slow autoregressive transformer predicts one semantic token for each audio frame. A smaller fast transformer predicts the remaining codec codebooks for that frame. The bundled neural codec then turns those tokens into a 44.1 kHz waveform.