Hey Maya wake word, version 1 (microWakeWord, int8 streaming)

README.md

@@ -0,0 +1,87 @@
+---
+license: apache-2.0
+language:
+- en
+tags:
+- wake-word
+- keyword-spotting
+- microwakeword
+- tflite
+---
+
+# Hey Maya — wake word, version 1
+
+"Hey Maya" is the phrase meant to start Maya, the Maya Foundation's assistant, on a person's phone, only
+when that person has turned the "Hey Maya" setting on (it is off by default). This is the small model
+that listens for the phrase. It runs on the phone; no audio needs to leave it.
+
+## The files
+
+| File | What it is |
+| --- | --- |
+| `hey_maya-v1.tflite` | The model: 62,304 bytes, SHA-256 `e95b7a1df30e0619dfa1f72134b0153d49f825cf4e63e656aa79ec0208da61bd`. A [microWakeWord](https://github.com/kahrendt/microWakeWord) streaming model that keeps its own state, quantised to int8 (microWakeWord's `stream_state_internal_quant` export). |
+| `training_parameters.yaml` | The training settings, as used. |
+| `tflite_streaming_roc.txt` | The test results below, as the training run wrote them. |
+
+## Input and output
+
+- **Features:** 16 kHz mono audio turned into 40-value frames by TensorFlow Lite Micro's audio front end
+  (the one microWakeWord trains with), one frame every 10 ms.
+- **Input:** `[1, 3, 40]` int8 (scale 0.10196078568696976, zero point −128): three frames, 30 ms, per call.
+- **Output:** `[1, 1]` uint8 (scale 1/256): how sure the model is that it just heard "Hey Maya".
+- **State:** the model keeps its memory between calls in TensorFlow Lite resource variables
+  (`VAR_HANDLE`, `READ_VARIABLE`, `ASSIGN_VARIABLE`, `CALL_ONCE`), so the interpreter must support them
+  and must be called in order, once per 30 ms.
+- **Deciding:** the test below averaged the last 5 outputs (150 ms, microWakeWord's default for this test)
+  and compared the average with the cutoff.
+
+## How well it works — on test audio, not on real people yet
+
+Measured by the training run on held-out audio. "Misses" is how often it did not wake on "Hey Maya" in
+synthetic test speech; "wakes by mistake" counts wakes on speech and music with no "Hey Maya" in it.
+
+| Cutoff | Misses "Hey Maya" | Wakes by mistake, per hour |
+| --- | --- | --- |
+| 0.99 (chosen for the app) | 15.7 % | 1.18 |
+| 0.98 | 13.9 % | 1.45 |
+| 0.97 | 13.1 % | 1.80 |
+| 0.96 | 12.8 % | 2.00 |
+
+AUC 0.795. **It has never heard a real voice:** every "Hey Maya" it learned from, and every one it was
+tested on, was synthetic speech (a different speech engine for the test than for training). Accuracy on
+real voices, in real rooms, is to be measured on phones before release and will differ from this table.
+Version 2 is to be trained on real recordings that people choose to share.
+
+## What it was trained on
+
+Only data the app may ship with (checked 2026-10-21):
+
+- **"Hey Maya" for training:** synthetic speech from the [Piper sample generator](https://github.com/rhasspy/piper-sample-generator) (MIT), using its 904-speaker LibriTTS-R voice model (LibriTTS-R, CC BY 4.0).
+- **"Hey Maya" for testing:** Phonon-31m English and its 37 English voices (Apache 2.0), a different engine, so the test is not the training voices again.
+- **Near misses** it must not wake on ("hey mia", "hey mayor", "maya" on its own, other assistants' names and more): the same Piper generator.
+- **Ordinary speech:** LibriSpeech dev and test sets (CC BY 4.0), with every utterance whose transcript says Maya, Maia, Maja, Mayer or Mayas removed.
+- **Noise and music:** MUSAN (CC BY 4.0), 80 % of the files for training and 20 % held out for the test.
+- **Room echo:** OpenSLR 28 RIRS_NOISES (Apache 2.0).
+
+Credits: LibriTTS-R (Koizumi et al., CC BY 4.0), LibriSpeech (Panayotov et al., CC BY 4.0), MUSAN (Snyder,
+Chen and Povey, CC BY 4.0), RIRS_NOISES (Ko et al., Apache 2.0), Phonon-31m (Apache 2.0), Piper sample
+generator (MIT), microWakeWord (Kevin Ahrendt, Apache 2.0).
+
+## How it was made
+
+microWakeWord (Apache 2.0) at commit `4665173cd35f1cff9a61e06fc427f124766c488e`, its "mixednet" model
+(`--pointwise_filters 64,64,64,64 --repeat_in_block 1,1,1,1 --mixconv_kernel_sizes '[5], [7,11], [9,15], [23]'
+--residual_connection 0,0,0,0 --first_conv_filters 32 --first_conv_kernel_size 5 --stride 3`), 20,000
+training steps, trained on 2026-10-21 on one rented GPU, then quantised to the streaming TensorFlow Lite
+model above.
+
+## Use and limits
+
+- For starting Maya when its owner has turned "Hey Maya" on. Not for telling who is speaking.
+- English, and the phrase "hey maya" only. "Maia" and "Maja" sound the same and were left out of the near
+  misses on purpose; expect it to wake on them too.
+- It sometimes wakes by mistake and sometimes misses; a tap on Maya's pill always works as well.
+
+## Licence
+
+The model: Apache 2.0. The training data keeps its own licences, credited above.

hey_maya-v1.tflite

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e95b7a1df30e0619dfa1f72134b0153d49f825cf4e63e656aa79ec0208da61bd
+size 62304

tflite_streaming_roc.txt

@@ -0,0 +1,6 @@
+AUC 0.79460
+Cutoff 1.00: frr=1.0000; faph=0.000
+Cutoff 0.99: frr=0.1568; faph=1.177
+Cutoff 0.98: frr=0.1392; faph=1.454
+Cutoff 0.97: frr=0.1311; faph=1.801
+Cutoff 0.96: frr=0.1284; faph=2.000

training_parameters.yaml

@@ -0,0 +1,55 @@
+batch_size: 128
+clip_duration_ms: 1500
+eval_step_interval: 500
+features:
+- features_dir: /workspace/tm2/wake/features/wakeword
+  penalty_weight: 1.0
+  sampling_weight: 2.0
+  truncation_strategy: truncate_start
+  truth: true
+  type: mmap
+- features_dir: /workspace/tm2/wake/features/near
+  penalty_weight: 1.0
+  sampling_weight: 6.0
+  truncation_strategy: truncate_start
+  truth: false
+  type: mmap
+- features_dir: /workspace/tm2/wake/features/speech
+  penalty_weight: 1.0
+  sampling_weight: 10.0
+  truncation_strategy: random
+  truth: false
+  type: mmap
+- features_dir: /workspace/tm2/wake/features/noise
+  penalty_weight: 1.0
+  sampling_weight: 5.0
+  truncation_strategy: random
+  truth: false
+  type: mmap
+- features_dir: /workspace/tm2/wake/features/ambient
+  penalty_weight: 1.0
+  sampling_weight: 0.0
+  truncation_strategy: split
+  truth: false
+  type: mmap
+freq_mask_count:
+- 0
+freq_mask_max_size:
+- 0
+learning_rates:
+- 0.001
+maximization_metric: average_viable_recall
+minimization_metric: null
+negative_class_weight:
+- 20
+positive_class_weight:
+- 1
+target_minimization: 0.9
+time_mask_count:
+- 0
+time_mask_max_size:
+- 0
+train_dir: /workspace/tm2/wake/trained
+training_steps:
+- 20000
+window_step_ms: 10