bench: add Pocket TTS, DeepFilterNet3 and DTLN to the CI bench suite

[czoli1976]

Adds three CPU speech workloads to the bench suite.

Pocket TTS (Kyutai, 100M-param CPU TTS): the flow-LM streaming decode step (one 80 ms frame) and 25-token prefill, one Euler step of the latent flow net, and a 12-frame Mimi decoder chunk. This is an ONNX KV-cache streaming decode workload — ScatterND cache writes, full-window masked attention, per-frame GEMV — a shape the suite doesn't cover today, and precisely the path exercised by #2446/#2448. flow_lm_flow.onnx and mimi_decoder.onnx are unmodified from the community ONNX export (KevinAHM/pocket-tts-onnx, weights CC-BY-4.0 from kyutai/pocket-tts); flow_lm_main_static.onnx is that export's flow LM with its runtime-step cache slicing rewritten shape-static (arange-table slices → Range(0,s)+step; window reads → full-window, the causal mask is a key-pos≤query-pos compare so out-of-window keys self-mask). The rewrite is validated against onnxruntime on the original graph to ≤2e-6 across (step, seq, text) combinations, and a full tract generation pipeline built on these graphs produces audio matching onnxruntime (corr 0.99999). Random-input note: --allow-random-input casts the random f32 [0,1) to 0 for the i64 step/offset states — a valid stream position — and the graphs are full-window, so compute is position-independent. M-series reference: decode 22.5 ms, prefill25 84.5 ms, flow 1.0 ms, mimi chunk12 136 ms.

DeepFilterNet3 (Rikorose/DeepFilterNet, MIT/Apache-2.0): the three streaming submodels (encoder, ERB-mask decoder, deep-filter decoder), symbolic time pinned at export to 100 frames = 1 s of audio at the 10 ms hop. Depthwise-conv + GRU heavy. M-series reference: 7.6 / 7.3 / 6.7 ms.

DTLN (breizhn/DTLN, MIT): both LSTM stages of the dual-signal denoiser, one 8 ms frame per call with LSTM state threaded through graph I/O — the external-state pattern of issue #2157. model_2's free batch dim on outputs is pinned to 1 (tf2onnx export artifact; tract rejects the unpinned graph). M-series reference: 38 / 67 µs per frame.

Needs from a maintainer: mirror the model bundles to the S3 bucket. Archives ready to hand over — I can stage them as release assets on my fork or anywhere convenient:

sha256  64daef9d8dfe780d2840d671ee2f9fb930ca13ef455bc5438df51ee488b41a0a  pocket-tts-en.tar.gz  (173 MB)
sha256  47052090e4884e39c531785868080e4e3c3d7d9aefba29dbf88a69a287bf4d13  dfn3.tar.gz           (7.6 MB)
sha256  b89911856e36cd93430be59a594d9bb5b231c7737d4d52b0516e0e7de7f28388  dtln.tar.gz           (3.5 MB)
-> s3://tract-ci-builds/model/

Until they're mirrored, the Bench run on this PR will 404 on the fetch — happy to keep this as draft until then. (#2446, which the shape-static flow-LM graph needs to load, and #2448, which the Pocket TTS ScatterND cache writes exercise, are both merged.)

All nine entries validated end-to-end through tract bench-suite --manifest .travis/benches.toml against a pre-populated cache (66 metrics emitted).

🍍

🤖 Generated with Claude Code

[czoli1976]

@kali - do you think you can add these ? It is draft as would need you to add them to the S3 Bucket

[kali]

I think so, yes. Let's park it for a while (a week or so), I'm in the middle of moving out of s3 for the model zoo.

[czoli1976]

NAS? Best Regards Ckristian Zoli Email: ***@***.***

…

On Mon, 6 Jul 2026 at 12:36 Mathieu Poumeyrol ***@***.***> wrote: *kali* left a comment (sonos/tract#2449) <#2449 (comment)> I think so, yes. Let's park it for a while (a week or so), I'm in the middle of moving out of s3 for the model zoo. — Reply to this email directly, view it on GitHub <#2449?email_source=notifications&email_token=APL2Z6WKWN3US3BD2TD2BND5DOFL3A5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTIOBZGIZTEMBTGAYKM4TFMFZW63VGMF2XI2DPOKSWK5TFNZ2KYZTPN52GK4S7MNWGSY3L#issuecomment-4892320300>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/APL2Z6RA5XQKAOGWRJC5YCD5DOFL3AVCNFSNUABEKJSXA33TNF2G64TZHM4TSNJWGEZDCMZ3JFZXG5LFHM2DQMJXGY4TKOJRHGQXMAQ> . Triage notifications, keep track of coding agent tasks and review pull requests on the go with GitHub Mobile for iOS <https://github.com/notifications/mobile/ios/APL2Z6TRFA3LQRPZBHYHRRT5DOFL3A5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTIOBZGIZTEMBTGAYKM4TFMFZW63VGMF2XI2DPOKSWK5TFNZ2KUZTPN52GK4S7NFXXG> and Android <https://github.com/notifications/mobile/android/APL2Z6WGYBLEQY2A3CQDTWL5DOFL3A5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTIOBZGIZTEMBTGAYKM4TFMFZW63VGMF2XI2DPOKSWK5TFNZ2K4ZTPN52GK4S7MFXGI4TPNFSA>. Download it today! You are receiving this because you authored the thread.Message ID: ***@***.***>

[kali]

I honestly considered it, with the right caching, the CI would probably be fine. R2 is the current contender.