s6-ear2-nano
Passed its pre-registered gate. Every condition held. A 632 KB ternary end-of-turn detector: given the last 4 s of audio it says whether the speaker has finished their turn. On the 31,527 test clips of Smart Turn v3.2 it is right on 0.8993 (bars: β₯ 0.85 and β₯ the teacher β 0.08, the teacher measured at 0.9260 on the same clips through the same harness); at its validation threshold Οββ it would cut a speaker off on 0.0847 of the not-finished clips (bar β€ 0.20) while catching 0.8627 of the finished ones (bar β₯ 0.80); it decides in 2.39 ms per window in WASM (bar 5 ms) from a file of 631,936 bytes (bar 3 MB), and the browser build matches the native one to 1.2e-7 with the same decision on 20 of 20 clips. The gate is
docs/s6-ear2-prereg.mdΒ§9, frozen before any line of v2's code, with the harness checked against Pipecat's published number before the test was read. Numbers: BENCH.md, section s6-ear2-nano andruns/s6-ear2-nano/GATE.md.When to take which model (the sentence of the gate report). "Take v2 when size and speed matter: it gives up 2.7 points of accuracy against Smart Turn for a file of 0.63 MB instead of 8.7 MB and about 2.4 ms per decision in WASM instead of about 32 ms natively; take Smart Turn v3.2's file when those points matter more."
Know before you use it. On real conversational speech (LiveKit's eot-bench, 17,934 cuts, which the design never read) this model and its teacher are far weaker than on the test clips: AUC 0.7299 against 0.7838, and at Οββ this model answers "finished" on 0.5253 of the real ends of turn β about half pass. Its weakest languages are Vietnamese (0.7749), Marathi (0.7907) and Bengali (0.7950). It is a student of Smart Turn: it cannot be better than its teacher's labels. Most of its test set is synthetic speech. The official test set is partly in-sample for the design choices, though never for the weights. Try it with your microphone on the site's Listen page or in the Space; the demo runs a fallback timer next to the model, for the reason above (How to use it).
Its successor failed, so this stays the file to take. s6-ear3-nano is this graph trained again on real two-speaker conversations and on the silences the demo asks at. On eot-bench read 0.2 s into each pause, it reads 0.7197 where this model reads 0.7448 on the same cuts: it failed its pre-registered gate and is published as a result. It was designed after this model's eot-bench reading, and its label cross-check was amended after a measurement on five training sessions; its card discloses both.
A dedicated model, not a shared one. The first version, s6-ear-nano,
shared one ternary model between end of turn and keywords on top of a codec pretraining, and
failed three of its four gates (end of turn 0.5681). This version drops all of that: one task,
no pretraining, a dedicated turn graph (a patch projection, 4 s2 encoder blocks, Smart Turn's
attention pooling, a 2-layer head; 1,306,178 ternary parameters), and a front end built for the
job β 40 log-mel bands plus 3 prosody channels (log f0, voicing, energy) per 10 ms hop. It was
trained for 240 working minutes on one Apple M5 as a student of Smart Turn v3.2, once, seed 1.
- Demo: in the browser, with the microphone or a self-test without one: the site's Listen page and the Space maxgfr/s6-ear2-nano. The owner checked the microphone path on the site in Chrome on 17/11/2026; Safari, and the microphone inside the Hugging Face frame, were not checked. How the demo decides is in How to use it.
- Code: github.com/maxgfr/skycmd (100 % Rust; the model is
sky_infer::turn, the front endcrates/sky-infer/src/audio.rs) - All results, failures included: results page
- Technical report: maxgfr.github.io/skycmd/paper
- Pre-registration: docs/s6-ear2-prereg.md (the design in Β§2, the gate in Β§9, amendment 1, the result block)
- Data: its feature stores are not republished; they are rebuilt from the public Smart Turn v3.2 sets (Training data and licences, below)
- Every demo, with its verdict: the skycmd Space
The gate, as read
The 31,527 clips of pipecat-ai/smart-turn-data-v3.2-test, the ternary GGUF as exported
(runs/s6-ear2-nano/GATE.md; 95 % bootstrap intervals, 1,000 resamples by clip):
| gate | condition | measured | verdict |
|---|---|---|---|
| G1 | accuracy at Οβ.β β₯ 0.85 and β₯ T β 0.08 (T = Smart Turn v3.2 CPU, same harness) | 0.8993 [0.8962, 0.9027] against T = 0.9260 | holds |
| G2 | false cut-off at Οββ β€ 0.20 | 0.0847 [0.0800, 0.0888] | holds |
| G2 | END recall at Οββ β₯ 0.80 | 0.8627 [0.8577, 0.8680] | holds |
| G3 | WASM median β€ 5.0 ms per decision | 2.39 ms (p95 2.41 ms) | holds |
| G3 | front end β€ 0.10 ms per 10 ms hop in WASM | 0.0186 ms | holds |
| G3 | GGUF β€ 3 MB | 631,936 bytes | holds |
| G3 | WASM parity: max |Ξp| β€ 1e-4, same decision | 1.2e-7, 20 of 20 clips | holds |
At Οβ.β : precision 0.8776, recall 0.9271, F1 0.9017; AUC 0.9600 [0.9581, 0.9621], ECE 0.0160 [0.0138, 0.0191]. The silence timer (always "finished") scores 0.4979 on the same clips.
On the same test clips
| model | accuracy | file | per decision |
|---|---|---|---|
| s6-ear2-nano | 0.8993 [0.8962, 0.9027] | 631,936 bytes | 2.39 ms in WASM, 1.40 ms native |
| Smart Turn v3.2 CPU (int8, through the same harness) | 0.9260 | 8.7 MB | 31.7 ms native |
| Smart Turn v3.2 GPU (Pipecat's published number) | 0.9371 | ||
| v1's control (no Mimi pretraining, not published) | 0.7647 | ||
| s6-ear-nano (v1, failed its gate) | 0.5681 | 899 KB | |
| silence timer (always "finished") | 0.4979 | no parameters |
The harness was checked before the test was read: Smart Turn v3.2 CPU through it gives 0.9260 on
these clips, within Pipecat's reported 0.9263 Β± 0.003, and the timer gives the test's share of
finished clips, 0.4979 (runs/s6-ear2-harness/check.json).
By language (accuracy at Οβ.β
; the full tables are in GATE.md): German 0.9493, Korean 0.9449,
Dutch 0.9471, Portuguese 0.9421 on the strong side; English 0.9056 (the English human subsets
0.9187); Vietnamese 0.7749, Marathi 0.7907 and Bengali 0.7950 on the weak side. By dataset,
the weakest is mundo_1 at 0.7843.
On real conversations (eot-bench, reported, not gated)
LiveKit's eot-bench is real conversational speech, which the design never read. One decision per labelled pause, cut at its start; 17,934 cuts. These numbers are not comparable with LiveKit's FC@300 / FC@600, which come from a policy sweep:
| model | AUC | false cut-off | eot recall |
|---|---|---|---|
| s6-ear2-nano (at Οββ ) | 0.7299 | 0.2230 | 0.5253 |
| Smart Turn v3.2 CPU (at 0.5) | 0.7838 | 0.3678 | 0.7998 |
| timer | 0.5000 | 1.0000 | 1.0000 |
The pre-registration fixed the reading in advance: "does not carry over" is said only if the student's AUC is more than 0.10 below the teacher's. It is 0.054 below, so that sentence does not apply. Said plainly: both models are far weaker on these real conversational cuts than on the test clips, and at Οββ this model answers "finished" on 0.5253 of the real ends of turn β about half pass. On a real deployment, use Οββ only together with the fallback timer below.
How to use it
The input. 16 kHz mono PCM. The engine computes the front end itself β no external
extractor: a 25 ms Hann window, a 10 ms hop, 40 log-mel bands, and per hop the three prosody
channels (log f0, voicing, energy; pitch searched between 80 and 400 Hz by autocorrelation).
Each window is normalised per clip by Whisper's rule, and the model reads the last 4 s
(400 hops, 8 per position, 50 positions; a shorter clip gives fewer positions). One forward pass
gives the calibrated probability that the turn is finished, Ο(logit / temperature) with the
stored temperature 1.4948.
The two thresholds in the file. Both come from validation, never from test: Οβ.β = 0.5 on the calibrated probability β the accuracy reading (0.8993 on the test clips) β and Οββ = 0.7244, the highest threshold whose validation END recall is at least 0.85 β the reading for a voice agent that must not cut people off (false cut-off 0.0847, END recall 0.8627 on test). Answer "finished" when the probability reaches the threshold, else keep listening.
In the browser or Node (sky-web, the TurnDetector class; the module is built by
scripts/build-wasm.sh):
import { TurnDetector } from './pkg/sky_web.js'
const det = new TurnDetector(new Uint8Array(ggufBytes))
det.push(pcmChunk) // 16 kHz mono Float32Array, any chunk size
const p = det.probEnd() // calibrated probability, from the stream's last 4 s
const end = p >= det.tau85() // or det.tau50() for the accuracy reading
const q = det.probEndPcm(clip) // one decision on a whole clip; the stream is untouched
det.reset() // a new turn
The front end computes each hop once and keeps the last 400, so streaming and batch features
agree bit for bit; probEnd costs the model's forward pass only (2.39 ms in WASM).
Natively (sky-turn, crates/sky-infer/src/bin/sky-turn.rs):
cargo run --release -p sky-infer --bin sky-turn -- s6-ear2-nano.gguf --records clips.jsonl --out probs.jsonl
cargo run --release -p sky-infer --bin sky-turn -- s6-ear2-nano.gguf --bench 1000
--records reads the 16 kHz WAV named by each JSONL row and prints {id, logit, prob, positions} per item; --bench times decisions on a full 4 s window.
In a voice agent (the operating notes of the pre-registration, Β§10): run the model once when
the VAD reports a pause (Silero, or the engine's energy VAD), with Pipecat's stop_secs 0.2;
answer "finished" when the probability is at least the threshold (Οβ.β
for accuracy, Οββ
for 85 %
END recall on validation), else keep listening; a fallback timer ends the turn after a longer
silence (Pipecat's SmartTurnParams.stop_secs); if speech resumes, the turn's audio is kept and
the model runs again at the next pause.
How the demo decides (the Listen page and the
Space run the same code). An energy VAD
over 10 ms frames finds the speech and the pauses; the model is asked once after 0.2 s of silence
following speech, then again every 0.2 s while the pause lasts; it answers "finished" at or above
the threshold, Οββ
by default (Οβ.β
can be chosen); and a fallback timer ends the turn after
2.5 s of silence whatever the model says. The timer is there because of eot-bench: on real
conversations, at Οββ
the model catches only 0.5253 of the real ends of turn, about half
(runs/s6-ear2-nano/GATE.md,
eot-bench), so without the timer a finished speaker could wait for an answer that never comes.
The pause length and the timer can be changed on the page.
Latency and size
Apple M5, one thread, 1,000 decisions on a full 4 s window with the features already computed,
measured with no training running (runs/s6-ear2-nano/latency/):
| native | WASM (SIMD128) | |
|---|---|---|
| median per decision | 1.40 ms | 2.39 ms |
| p95 | 1.42 ms | 2.41 ms |
| front end per 10 ms hop | 0.0134 ms | 0.0186 ms |
Smart Turn v3.2 CPU through onnxruntime, one thread, in the same session: median 31.7 ms per
decision. File: s6-ear2-nano.gguf, 631,936 bytes.
Training data and licences
One run, 240 working minutes (305,700 steps, about 152 epochs), batch 128, seed 1, on an Apple M5, in a braked mode, so the budget counts working minutes. Ternary QAT over the last 15 %, EMA 0.999; the exported checkpoint is the EMA. The temperature and the two thresholds are fitted on validation and stored in the file; the test set was read once, at the end.
- Training and validation:
pipecat-ai/smart-turn-data-v3.2-train(revisionf8c95d28, 270,946 clips), CC-BY-4.0, 5 % validation by a hash of the clip id (seed 1234): 257,353 training, 13,593 validation. A clip is dropped if its id is a test id or its decoded PCM hashes to a test clip's; the filter dropped no row (the sentence-level overlap cannot be checked,spoken_textbeing empty, as for Smart Turn's own numbers). - Test (gated):
pipecat-ai/smart-turn-data-v3.2-test, 31,527 clips, CC-BY-4.0 β Smart Turn v3.2's own official held-out set. Never trained on, never used for a threshold. - Second test (reported only):
livekit/eot-bench-data, CC-BY-4.0, real conversational speech, one decision per labelled pause; never trained on. - The teacher: Smart Turn v3.2
(
smart-turn-v3.2-cpu.onnx, revisionf766f81d), BSD-2-Clause, credited: its logit, its pooled embedding and its encoder states are the training targets on the training and validation clips only. Being a student, this model cannot be better than its teacher's labels. The teacher is never redistributed.
No audio is redistributed with the model: only the GGUF. Every licence, with the revision of each source, is in DATA_LICENSES.md.
Disclosure (Β§11 of the pre-registration)
This version was designed after v1's test results, and its design study read measurements made on 3,000 clips of the same test set: v1's two models per source, Smart Turn per source, Smart Turn with 8, 4, 3, 2 and 1 s of context, and the trailing silence of 1,500 clips. The 4 s window and the data change came from those readings, so the official test set is partly in-sample for the design choices, though never for the weights: no test clip is trained on, and no threshold is chosen on test. eot-bench, which the design never read, is the cleaner check, and it is reported above. The run happened under the automatic brake of the machine, and CPU work of other models ran during the pilot and part of the training (allowed by amendment 1; budgets count working minutes); the pilot and the training were built from commits that differ by an unrelated option, off by default, and nothing in the end-of-turn training path changed between them. The latencies were measured with no training running.
Files
s6-ear2-nano.gguf: ternary weights, the window and prosody configuration, the temperature (1.4948) and the two validation thresholds (Οβ.β 0.5, Οββ 0.7244),skycmd-turnarchitecture. 631,936 bytes, sha256204354653fd864708c3206f486f255d3fedff81956fd7b82aae560185a9be783.
- Downloads last month
- 51
We're not able to determine the quantization variants.