wakeforge 0.4.2a5__tar.gz → 0.4.2a7__tar.gz

{wakeforge-0.4.2a5 → wakeforge-0.4.2a7}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: wakeforge
-Version: 0.4.2a5
+Version: 0.4.2a7
 Summary: Full training and research suite for wake words — MFCC to HuBERT, all exported to ONNX
 Author-email: JarbasAi <jarbasai@mailfence.com>
 License: Apache-2.0
@@ -89,6 +89,13 @@ the result is safe to ship — see
 deployable wake word actually looks like in false-accept and false-reject
 numbers, and how much data each tier needs to get there.
 
+The first measured numbers show why. The first two models that wakeforge
+trained scored between 2397 and 5120 false accepts per hour on speech, noise
+and music at threshold 0.5, against a bar of 1. They accepted 70 % of the
+near-miss phrases, against a bar of 5 %. The training negatives were the cause.
+The rows, the commands and the cause are in
+[`docs/guides/expectations.md`](docs/guides/expectations.md#the-measured-baseline-two-rejected-models).
+
 ## What is a wake word?
 
 A short phrase ("hey jarvis", "computer", "alexa") that a device listens for

{wakeforge-0.4.2a5 → wakeforge-0.4.2a7}/README.md

@@ -17,6 +17,13 @@ the result is safe to ship — see
 deployable wake word actually looks like in false-accept and false-reject
 numbers, and how much data each tier needs to get there.
 
+The first measured numbers show why. The first two models that wakeforge
+trained scored between 2397 and 5120 false accepts per hour on speech, noise
+and music at threshold 0.5, against a bar of 1. They accepted 70 % of the
+near-miss phrases, against a bar of 5 %. The training negatives were the cause.
+The rows, the commands and the cause are in
+[`docs/guides/expectations.md`](docs/guides/expectations.md#the-measured-baseline-two-rejected-models).
+
 ## What is a wake word?
 
 A short phrase ("hey jarvis", "computer", "alexa") that a device listens for

{wakeforge-0.4.2a5 → wakeforge-0.4.2a7}/wakeforge.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: wakeforge
-Version: 0.4.2a5
+Version: 0.4.2a7
 Summary: Full training and research suite for wake words — MFCC to HuBERT, all exported to ONNX
 Author-email: JarbasAi <jarbasai@mailfence.com>
 License: Apache-2.0
@@ -89,6 +89,13 @@ the result is safe to ship — see
 deployable wake word actually looks like in false-accept and false-reject
 numbers, and how much data each tier needs to get there.
 
+The first measured numbers show why. The first two models that wakeforge
+trained scored between 2397 and 5120 false accepts per hour on speech, noise
+and music at threshold 0.5, against a bar of 1. They accepted 70 % of the
+near-miss phrases, against a bar of 5 %. The training negatives were the cause.
+The rows, the commands and the cause are in
+[`docs/guides/expectations.md`](docs/guides/expectations.md#the-measured-baseline-two-rejected-models).
+
 ## What is a wake word?
 
 A short phrase ("hey jarvis", "computer", "alexa") that a device listens for

{wakeforge-0.4.2a5 → wakeforge-0.4.2a7}/ww_trainer/version.py

@@ -2,7 +2,7 @@
 VERSION_MAJOR = 0
 VERSION_MINOR = 4
 VERSION_BUILD = 2
-VERSION_ALPHA = 5
+VERSION_ALPHA = 7
 # END_VERSION_BLOCK
 
 __version__ = f"{VERSION_MAJOR}.{VERSION_MINOR}.{VERSION_BUILD}" + (f"a{VERSION_ALPHA}" if VERSION_ALPHA else "")