Skip to content

Automation foundation model for tiny devices: 2-bit, 8-29 MB, tool calls, ASR, structured extraction and embeddings on phones, wearables, smart homes, robots, cars and microcontrollers.

fix: update version to 3.2.0 and improve audio length handling in Whistle

HenryNdubuaku24d5765dc2d2
diff --git a/README.md b/README.md
index 1c4bab7..db4b990 100644
--- a/README.md
+++ b/README.md
@@ -48,7 +48,7 @@ Every turn returns one JSON object with `function_calls`, the model's `reasoning
![One engine, three ways to load it](assets/whistle.svg)
-Whistle is our speech-to-text model, one 16.9 MB file on the CPU: 16 kHz mono audio, up to 30 seconds in one pass, in English, German, French, Spanish, Italian, Dutch and Polish. It shares Needle's `.cact` container, its quantisation and its C++ engine, so the two are one runtime, and its decoder is laddered the same way, with `--audio-depth N` running the N-layer rung of the same weights.
+Whistle is our speech-to-text model, one 16.9 MB file on the CPU: 16 kHz mono audio of any length, one pass up to 30 seconds and windowed beyond it, in English, German, French, Spanish, Italian, Dutch and Polish. It shares Needle's `.cact` container, its quantisation and its C++ engine, so the two are one runtime, and its decoder is laddered the same way, with `--audio-depth N` running the N-layer rung of the same weights.
```python
import needle
diff --git a/needle/__init__.py b/needle/__init__.py
index 5993c45..b940c08 100644
--- a/needle/__init__.py
+++ b/needle/__init__.py
@@ -13,7 +13,7 @@ from ._telemetry import track as _track
from ._worker import FineTuneWorker
from .agent.whistle import Whistle, stream, transcribe
-__version__ = "3.1.0"
+__version__ = "3.2.0"
__all__ = ["Needle", "Whistle", "ExtractionValidationError", "tool", "Field", "extract",
"transcribe", "stream", "__version__"]
diff --git a/needle/agent/fetch.py b/needle/agent/fetch.py
index 33d3491..c4627b3 100644
--- a/needle/agent/fetch.py
+++ b/needle/agent/fetch.py
@@ -13,7 +13,7 @@ ENGINE_REPOS = {
}
ENGINE_VERSIONS = {
2: "2.0.4",
- 3: "3.2.0",
+ 3: "3.3.0",
WHISTLE: "2.0.0",
}
diff --git a/needle/agent/whistle.py b/needle/agent/whistle.py
index 109d883..bd189f9 100644
--- a/needle/agent/whistle.py
+++ b/needle/agent/whistle.py
@@ -98,7 +98,7 @@ def _samples(audio):
class Whistle:
"""Speech-to-text in English, German, French, Spanish, Italian, Dutch and Polish.
- Audio is a WAV file path or 16 kHz mono float samples in [-1, 1], at most 30 s.
+ Audio is a WAV file path or 16 kHz mono float samples in [-1, 1], of any length.
One model is loaded per process; it is not thread-safe.
"""
@@ -181,7 +181,7 @@ def _shared_model(weights):
def transcribe(audio, language=None, keywords=None, word_timestamps=False, weights=None) -> dict:
"""Speech to text, on the model this process already has loaded.
- audio is a WAV file path or 16 kHz mono float samples in [-1, 1], at most 30 s.
+ audio is a WAV file path or 16 kHz mono float samples in [-1, 1], of any length.
Returns {"text", "language", "ttft_ms", "decode_tps"}, plus "words" with start,
end and probability when word_timestamps is set. language is one of LANGUAGES,
or None to detect it; keywords are words or phrases for keyword biasing.
@@ -220,7 +220,6 @@ COMPARE_INSTALL = 'pip install "cactus-needle[mic,compare]"'
CLEAR_ABOVE = "\x1b[1A\x1b[2K"
DIM = "\x1b[90m"
PLAIN = "\x1b[0m"
-LIMIT_SECONDS = 30
def audio_path(rest):
@@ -257,10 +256,7 @@ def record():
except sounddevice.PortAudioError as error:
raise RuntimeError(f"no microphone: {error}") from None
print(CLEAR_ABOVE, end="", flush=True)
- heard = numpy.concatenate(chunks)[:, 0] if chunks else numpy.zeros(0, numpy.float32)
- audio = heard[:LIMIT_SECONDS * rate]
- if len(audio) < len(heard):
- print(f" keeping the first {LIMIT_SECONDS} s")
+ audio = numpy.concatenate(chunks)[:, 0] if chunks else numpy.zeros(0, numpy.float32)
return audio if rate == SAMPLE_RATE else soxr.resample(audio, rate, SAMPLE_RATE, quality="HQ")
diff --git a/pyproject.toml b/pyproject.toml
index 82a4dec..97ab73e 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -1,6 +1,6 @@
[project]
name = "cactus-needle"
-version = "3.1.0"
+version = "3.2.0"
description = "A 14MB foundation tool-calling model for tiny devices: inference, LoRA finetuning, and build."
requires-python = ">=3.9"
readme = "README.md"
diff --git a/tests/test_whistle.py b/tests/test_whistle.py
index 1b85df7..78bb467 100644
--- a/tests/test_whistle.py
+++ b/tests/test_whistle.py
@@ -206,14 +206,16 @@ def test_transcribe_returns_text_language_and_timed_words(tmp_path):
@requires_whistle
-def test_embed_and_the_30_second_limit():
+def test_embed_keeps_its_window_and_transcribe_takes_long_audio():
from needle import Whistle
whistle = Whistle()
embedding = whistle.embed(_tone(1))
assert len(embedding) > 0 and all(isinstance(v, float) for v in embedding[:8])
with pytest.raises(RuntimeError, match="30 s"):
- whistle.transcribe([0.0] * (30 * 16000 + 1))
+ whistle.embed([0.0] * (30 * 16000 + 1))
+ long = whistle.transcribe([0.0] * (31 * 16000), word_timestamps=True)
+ assert long["text"] == "" and long["language"] == "" and long["words"] == []
@requires_whistle
@@ -265,7 +267,7 @@ class _Microphone:
return Stream()
-def test_record_resamples_the_microphone_to_16_khz_and_keeps_30_s(monkeypatch, capsys):
+def test_record_resamples_the_microphone_to_16_khz_and_keeps_everything(monkeypatch, capsys):
numpy = pytest.importorskip("numpy")
pytest.importorskip("soxr")
from needle.agent.whistle import record
@@ -278,8 +280,7 @@ def test_record_resamples_the_microphone_to_16_khz_and_keeps_30_s(monkeypatch, c
assert audio.dtype == numpy.float32 and len(audio) == 16000 and abs(float(audio[8000]) - 0.25) < 0.01
assert "recording, Enter to stop" in capsys.readouterr().out
monkeypatch.setitem(sys.modules, "sounddevice", _Microphone(16000, 31))
- assert len(record()) == 30 * 16000
- assert "keeping the first 30 s" in capsys.readouterr().out
+ assert len(record()) == 31 * 16000
def test_record_says_what_is_missing(monkeypatch):