Post

Hello! Olá! I didn't have any time yesterday to write this post! Sorry! Anyway, on October 29th Fish Audio quietly uploaded new Drama 3 weights. I patched my Studio to run them. This update also fixes phoneme words, stops the word check from rejecting good takes, and decodes about 12% faster.

Press Update in Pinokio. If you run the full model, Update downloads the new weights (39.6 GB), so keep that much disk free. The old files are deleted right after, and if the download stops, the Studio keeps working on the old weights until you run it again.

Fish's new weights

• Fish swapped the files with no announcement: 909 of 912 tensors changed, while the codec and tokenizer stayed the same.
• tts-bench found they pace like Fish's hosted app on English, within 5%, where the old weights ran 19-38% fast. A Spanish take that used to run to the frame limit now ends normally. Meanwhile, I found the new weights to realiably give better audio.
• All 108 test takes here passed the word check (12 lines, 3 voices, 3 seeds). The pause fix and stall guard needed no changes.
• I refit the list of experts kept in VRAM for the new weights: 72-79 frames/s, up from 65-70, on an RTX 5090. The takes come out the same.

Word check: fewer false alarms
Good takes were being remade because of how Whisper writes things. On 508 good recorded takes, 471 now pass, up from 413.

• Numbers said the way people say them: "3.30" for "three thirty", "$2.5 million", "555-0142" read digit by digit, "a hundred and seventy one".
• Units such as "5 km" and "250 ml".
• Whisper's labels for sounds, like "(coughing)" or "[BLANK_AUDIO]".
• Names Whisper writes in capitals, like "QUEN" for "Qwen".

Takes missing a word are caught exactly as often as before. One trade-off: a take that says one wrong number now passes more often. The check was built to catch skipped spans and endings, not single wrong words.

Phoneme words

• A script ending on phoneme words no longer has them cut off.
• A take that says a phoneme word no longer scores worse than one that skips it. This was remaking good takes and keeping bad ones.
• A take that leaves out a phoneme word is now made again in Quality Mode.
• CMU Arpabet works better than IPA with this model, and the Studio now says so.

Smaller fixes

• The Space's limit of 60 takes per 10 minutes is gone. It made sense for a shared demo, not for your own GPU.
• The server now stops if the shell that started it dies, so it can't sit there holding 20+ GB of VRAM. Tested on Linux; the Windows path is untested.
• If the voice cache fails, the take goes on and reads the whole prompt instead of failing.
• MLX's memory cache is capped at 2 GiB. Checked on Linux, not yet on a Mac.

On a Mac: the MLX models are still conversions of Fish's original weights, because no conversion of the new ones exists yet. The Small GGUF is unchanged too.

Thanks to the projects this builds on: 5uck1ess/tts-bench for finding the weight change, maralski/drama-studio for the recorded takes and word-check ideas, latent-variable/shipshow for the number and name fixes and the MLX cache cap, plotarmordev/drama-tensorfold and ashhart/TensorFold, and tildaslashalef/nuclis for the cue measurements.