mashbox-labs/maya1-TPU

maya1 on Google Cloud TPU

Experimental

Runs with mashbox run maya1 in the mashbox package, release 0.1.15.

Read our guide: run maya1 on TPU

Maya Research's open text-to-speech model: a voice description and the text to say in, 24 kHz speech out, on one TPU chip, with the SNAC 24 kHz decoder it speaks through. Emotion tags such as and go inline in the text. Made by Maya Research, ported to TPU by MASHBOX Labs.

• 0.266 to 0.269 s: first audio, one request, decoded while it is generated (programs compiled), on TPU v6e-1
• 0.536 to 0.545 s: compute per second of audio, one request at a time, whole take (programs compiled), on TPU v6e-1
• 4.95 s of audio per second: 8 requests at once (programs compiled), on TPU v6e-1
• 11.33 GB: chip memory peak (8 slots, 3,072 tokens each), on TPU v6e-1
• 0.450 to 0.452 s: first audio, one request, decoded while it is generated (programs compiled), on TPU v5e-1
• 0.903 to 0.917 s: compute per second of audio, one request at a time, whole take (programs compiled), on TPU v5e-1
• 4.65 s of audio per second: 8 requests at once (programs compiled), on TPU v5e-1
• 11.29 GB: chip memory peak (8 slots, 3,072 tokens each), on TPU v5e-1
• maya1-BF16.gguf (6.60 GB, 16.00 bits per weight): mean KL 0.00255, top-1 agreement 91.4% vs the float32 reference (38 prompts, 15,392 codes)

Quickstart

On a Google Cloud TPU VM (the setup guide gets you there from a new Google account; coding agents can follow AGENTS.md):

python3 -m pip install uv && python3 -m uv tool install --python 3.12 "mashbox[tpu]>=0.1.15" --with-executables-from huggingface_hub
~/.local/bin/mashbox run maya1 --text "Hello there. <laugh> It works." --voice alloy --out hello.wav

Results

Chip	Measure	Result	Baseline
TPU v6e-1	First audio, one request, decoded while it is generated (programs compiled)	0.266 to 0.269 s	n/a
TPU v6e-1	Compute per second of audio, one request at a time, whole take (programs compiled)	0.536 to 0.545 s	n/a
TPU v6e-1	8 requests at once (programs compiled)	4.95 s of audio per second	n/a
TPU v6e-1	Chip memory peak (8 slots, 3,072 tokens each)	11.33 GB	n/a
TPU v5e-1	First audio, one request, decoded while it is generated (programs compiled)	0.450 to 0.452 s	n/a
TPU v5e-1	Compute per second of audio, one request at a time, whole take (programs compiled)	0.903 to 0.917 s	n/a
TPU v5e-1	8 requests at once (programs compiled)	4.65 s of audio per second	n/a
TPU v5e-1	Chip memory peak (8 slots, 3,072 tokens each)	11.29 GB	n/a

Files

File	Format	Size (GB)	Bits per weight	Quality vs full precision	Speed
maya1-BF16.gguf	GGUF	6.60	16.00	mean KL 0.00255, top-1 agreement 91.4% vs the float32 reference (38 prompts, 15,392 codes)	first audio 0.27 s, 0.54 s of compute per second of audio for 1 request (TPU v6e-1)
snac/snac_24khz_decoder.npz	NumPy arrays (npz)	0.05	n/a	109.9 dB SNR or more vs the official SNAC decoder in float32 (38 prompts)	n/a

Chips

Chip	Status	Note
TPU v5e	Supported	one chip (v5e-1)
TPU v6e	Supported	one chip (v6e-1)
TPU v5p slice (v5p-8)	Not tested yet
TPU v6e slice (v6e-4)	Not tested yet

Kernels

This model runs on plain JAX. It ships no compiled kernels.

Links

Code	mashbox-labs/mashbox
Guide	mashbox.gg/docs/models/maya1
Original model	maya-research/maya1 by Maya Research
Setup from a new Google account	mashbox.gg/docs/google-cloud
Companies	hello@mashbox.gg

Licence

Weights: Apache-2.0 licence, from maya-research/maya1@b1b9b38c; hubertsiuzdak/snac_24khz@d73ad176: MIT; code: Apache-2.0 (MASHBOX Labs). snac/LICENSE in this repo is the licence file of hubertsiuzdak/snac@8f79a718, unchanged. Credit: hubertsiuzdak/snac@8f79a718, MIT, code (the SNAC decoder in plain JAX follows its decode path); its LICENSE ships unchanged in the package as LICENSE-SNAC beside that code.