remove turn-bench-submission/contract.py
Browse files
turn-bench-submission/contract.py
DELETED
|
@@ -1,44 +0,0 @@
|
|
| 1 |
-
"""Turn-taking benchmark submission interface. See README.md for usage.
|
| 2 |
-
|
| 3 |
-
Implement ONE track:
|
| 4 |
-
- DiscriminativeModel — you declare the floor per timestep; we run it streaming.
|
| 5 |
-
- GenerativeModel — you produce audio; we VAD it into the floor track.
|
| 6 |
-
"""
|
| 7 |
-
|
| 8 |
-
from typing import Literal, Protocol, runtime_checkable
|
| 9 |
-
|
| 10 |
-
import numpy as np
|
| 11 |
-
from numpy.typing import NDArray
|
| 12 |
-
|
| 13 |
-
SAMPLE_RATE = 48_000 # benchmark feed rate: corpus + model input. (Generative output may be any rate.)
|
| 14 |
-
|
| 15 |
-
AudioChunk = NDArray[np.float32] # 1-D mono float32, shape (n_samples,), in [-1, 1]
|
| 16 |
-
# `subject` is the channel whose floor you predict; `other` is the conversation
|
| 17 |
-
# partner. Subject is a role, not a fixed speaker — the runner puts each speaker
|
| 18 |
-
# in the subject slot in turn.
|
| 19 |
-
FloorBit = Literal[0, 1] # 1 = subject holds the floor, 0 = subject does not
|
| 20 |
-
|
| 21 |
-
|
| 22 |
-
@runtime_checkable
|
| 23 |
-
class DiscriminativeModel(Protocol):
|
| 24 |
-
input_sample_rate: int # runner resamples to this before every call
|
| 25 |
-
|
| 26 |
-
def reset(self) -> None: # clear ALL streaming state; called before each pass
|
| 27 |
-
... # over a conversation (once per subject slot)
|
| 28 |
-
|
| 29 |
-
def step(self, subject_audio: AudioChunk, other_audio: AudioChunk) -> FloorBit: ...
|
| 30 |
-
|
| 31 |
-
|
| 32 |
-
@runtime_checkable
|
| 33 |
-
class GenerativeModel(Protocol):
|
| 34 |
-
output_sample_rate: (
|
| 35 |
-
int # the rate your generate() output is at — any value; just report it
|
| 36 |
-
)
|
| 37 |
-
|
| 38 |
-
def reset(self) -> None: # clear ALL streaming state; called before each pass
|
| 39 |
-
... # over a conversation (once per subject slot)
|
| 40 |
-
|
| 41 |
-
def generate(self, subject_audio: AudioChunk) -> AudioChunk:
|
| 42 |
-
# Return your model's raw audio response covering the same DURATION as subject_audio.
|
| 43 |
-
# We assert len(output) / output_sample_rate == len(subject_audio) / SAMPLE_RATE.
|
| 44 |
-
...
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|