Internal briefing · Jarvis voice
KittenTTS v0.8 runs in stock Sherpa and stays inside the memory ceiling, but its fastest useful checkpoint reaches audible PCM in 725 ms and cancels in 745 ms. Correctly pinned Piper Plus v1.13.0 is much faster and smaller than Kokoro, but its canonical English-capable voice scores 18.5% WER, fails the one-word prompt, and streams only at sentence boundaries.
5 September 2026
Intel i7-1255U, Fedora Linux. Three fresh processes and five warm turns per prompt.
Verdict
All four Kitten checkpoints execute through Jarvis's pinned Sherpa 1.13.6 runtime. No Sherpa patch or second inference engine is needed. Nano FP32 is the best trade: 0.13 RTF, 433 MiB peak RSS, a 61 MiB prepared model package, and 2.2% WER.
Nano FP32 misses the 250 ms warm-audio target by 475 ms and takes 745 ms to return after cancellation is requested. Mini, Micro, and Nano INT8 are slower. INT8 is also slower and heavier than Nano FP32 on this CPU.
The canonical Piper Plus graph needs seven inputs, including language, prosody, and speaker-embedding tensors. Sherpa's standard Piper/VITS runner supplies only the standard inputs plus optional speaker or language IDs. Supporting it there would require a new frontend and model contract, not a small metadata patch.
The pinned v1.13.0 C++ runtime reaches audible PCM in 117 ms warm, peaks at 303 MiB, and releases to 64 MiB without a Python TTS environment. The public Tsukuyomi checkpoint scores 18.5% WER and produces no recognized word for “Done.” Cancellation returns after 272 ms, but only after emitting two sentence chunks containing 14.5 seconds of audio.
Measured matrix
Latency values are medians. Peak, resident idle, and released RSS are maxima so the repeated-turn process is not hidden by the two cold-only processes. CPU is percent of one logical core; about 200% means both pinned cores were busy.
| Engine | Warm audible | Cold audible | RTF | CPU | Peak RSS | Idle RSS | Released RSS | Cancel | Chunks |
|---|---|---|---|---|---|---|---|---|---|
| Kokoro current | 1,924 ms | 4,101 ms | 1.07 | 199% | 845 MiB | 822 MiB | 179 MiB | 3,978 ms | 4 |
| Pocket official FP32 | 236 ms | 4,288 ms | 0.30 | 170% | 1,188 MiB | 1,163 MiB | 931 MiB | 8 ms | 107 |
| Pocket official INT8 | 160 ms | 4,692 ms | 0.25 | 180% | 1,133 MiB | 1,003 MiB | 724 MiB | 9 ms | 106 |
| Pocket Sherpa January | 3,053 ms | 5,374 ms | 0.75 | 199% | 669 MiB | 655 MiB | 68 MiB | 1,902 ms | 8 |
| Pocket current native prototype | 1,747 ms | 3,361 ms | 0.94 | 199% | 579 MiB | 579 MiB | 68 MiB | 957 ms | 9 |
| Supertonic 3 Sherpa INT8 | 3,652 ms | 4,630 ms | 0.41 | 199% | 430 MiB | 400 MiB | 73 MiB | 8,039 ms | 1 |
| Kitten v0.8 Mini Sherpa | 1,572 ms | 2,283 ms | 0.65 | 199% | 687 MiB | 659 MiB | 71 MiB | 3,340 ms | 4 |
| Kitten v0.8 Micro Sherpa | 844 ms | 1,609 ms | 0.27 | 199% | 547 MiB | 504 MiB | 72 MiB | 1,880 ms | 4 |
| Kitten v0.8 Nano FP32 Sherpa | 725 ms | 1,251 ms | 0.13 | 199% | 433 MiB | 389 MiB | 71 MiB | 745 ms | 4 |
| Kitten v0.8 Nano INT8 Sherpa | 1,201 ms | 1,829 ms | 0.30 | 199% | 553 MiB | 510 MiB | 71 MiB | 1,707 ms | 4 |
| Piper Plus v1.13 native C++ | 117 ms | 1,190 ms | 0.04 | 199% | 303 MiB | 303 MiB | 64 MiB | 272 ms | 4 |
The four Kitten variants allocate ONNX arena memory during the first two variable-length cycles, then settle. Piper Plus received a separate 20-turn soak: after its first allocator warm-up, RSS moved from 292.34 MiB to 294.58 MiB, a 2.24 MiB increase over the next 19 typical turns. Its 25-word response arrives in four sentence chunks, but the first chunk itself contains about 1.58 seconds of audio. The long prompt's first chunk contains about 4.36 seconds.
Intelligibility
Jarvis Parakeet transcribed the exact saved output after resampling to 16 kHz, without hotwords. WER is an objective intelligibility proxy, not a naturalness score. Human listening is still required for timbre, prosody, and preference.
| Engine | WER | Typical prompt | Interpretation |
|---|---|---|---|
| Kokoro current | 6.5% | Perfect | Production baseline |
| Pocket official FP32 | 4.3% | Perfect | Reference only |
| Pocket official INT8 | 5.4% | Perfect | Reference only |
| Pocket current native | 5.4% | Perfect | Clear but slow |
| Supertonic 3 | 6.5% | Perfect | Clear but non-streaming |
| Kitten Mini | 2.2% | Perfect | Best WER, slowest Kitten |
| Kitten Micro | 5.4% | 8.0% | Slower and less accurate than Nano FP32 |
| Kitten Nano FP32 | 2.2% | Perfect | Best Kitten trade |
| Kitten Nano INT8 | 2.2% | Perfect | Quantization regresses speed and memory |
| Piper Plus v1.13 native, published settings | 18.5% | 20.0% | “Done.” yields no ASR word; Japanese-accented checkpoint needs human review |
Method and compatibility
Each engine ran in a fresh process pinned to CPUs 0 and 2 with two inference threads. Each first process ran five warm repetitions of the same one-word, 25-word, and 66-word prompts. Two more processes supplied cold-start samples. Model files were local and eligible for the OS page cache. Physical DAC and speaker buffering were excluded.
Warm first audio is request-to-first audible PCM, where audible means the first 10 ms window at or above −45 dBFS. Cancellation was requested 250 ms into the long prompt and ended when synthesis returned. RSS came from Linux /proc/self/status. Released RSS was measured after deleting model references, garbage collection, and malloc_trim.
Kitten used stock Sherpa 1.13.6 and Sherpa's upstream v0.8 metadata-packaging script. The decision-grade Piper run used the official v1.13.0 tag at 7d3aa34, built as its native C API library, with ONNX Runtime 1.20 and the release's own C++ English dictionary. The model was tsukuyomi-chan-6lang-fp16.onnx with speaker 0, automatic language detection, noise_scale=0.667, length_scale=1.5, and noise_w=0.8. A separate explicit-English run scored worse at 25.0% WER.
The benchmark host process adds about 23 MiB of Python and NumPy baseline memory, but all Piper phonemization and inference in the reported row run in the pinned C++ library; there is no Python TTS dependency. The earlier moving-branch Python result is superseded because it used package 2.0.0 and the config's length_scale=1.0, not the checkpoint's published invocation.
Raw files are under /home/abstergo/Documents/Jarvis-pockettts-benchmark/benchmarks/tts-candidates. Eight focused benchmark tests pass, all JSON parses, and git diff --check is clean.
Sources checked 5 September 2026
KittenML KittenTTS documents the v0.8 checkpoints. Sherpa issue 3196 tracks v0.8 ONNX2 compatibility.
Piper Plus v1.13.0 is the pinned native runtime. The Tsukuyomi model card supplies the canonical public checkpoint and the published inference settings used here.
Sherpa PocketTTS, current Pocket compatibility issue, and Sherpa Supertonic support the earlier rows.