A demo for POWSM's four tasks

.github/workflows/s2t_pr_demo.yml

@@ -0,0 +1,31 @@
+# Demos/s2t_pr_demo.ipynb, every Sunday. The badge on this workflow is what
+# the README shows beside that notebook, which is why it has a file of its own:
+# GitHub's badge is per workflow, and cannot show one job of a matrix.
+
+name: s2t_pr_demo
+
+on:
+  schedule:
+    # Sundays, 05:00 UTC
+    - cron: "0 5 * * 0"
+  workflow_dispatch:
+  pull_request:
+    paths:
+      - "Demos/s2t_pr_demo.ipynb"
+      - "tools/**"
+      - ".github/workflows/s2t_pr_demo.yml"
+      - ".github/workflows/_run_notebook.yml"
+
+permissions:
+  contents: read
+
+jobs:
+  run:
+    uses: ./.github/workflows/_run_notebook.yml
+    with:
+      notebook: Demos/s2t_pr_demo.ipynb
+      # POWSM's <g2p> and <p2g> reach the page through espnet2.bin.demo, and
+      # this notebook pins the release that carries them (espnet#6791 cuts
+      # it). Until it is on PyPI the job says so and skips. Delete this line
+      # when 202610.post2 is published.
+      allow_unreleased_pin: true

Demos/README.md

@@ -40,6 +40,7 @@ the answer now.
 | [`asr_streaming_demo.ipynb`](asr_streaming_demo.ipynb) | [![asr_streaming_demo](https://github.com/espnet/notebook/actions/workflows/asr_streaming_demo.yml/badge.svg)](https://github.com/espnet/notebook/actions/workflows/asr_streaming_demo.yml) | Watch the words appear while the audio is still arriving |
 | [`st_demo.ipynb`](st_demo.ipynb) | [![st_demo](https://github.com/espnet/notebook/actions/workflows/st_demo.yml/badge.svg)](https://github.com/espnet/notebook/actions/workflows/st_demo.yml) | Translate English speech into German, French and Chinese — the same model |
 | [`s2t_align_demo.ipynb`](s2t_align_demo.ipynb) | [![s2t_align_demo](https://github.com/espnet/notebook/actions/workflows/s2t_align_demo.yml/badge.svg)](https://github.com/espnet/notebook/actions/workflows/s2t_align_demo.yml) | Line text up with the audio it was said in, and score how well they agree |
+| [`s2t_pr_demo.ipynb`](s2t_pr_demo.ipynb) | [![s2t_pr_demo](https://github.com/espnet/notebook/actions/workflows/s2t_pr_demo.yml/badge.svg)](https://github.com/espnet/notebook/actions/workflows/s2t_pr_demo.yml) | Hear the phones, and cross between phones and words with POWSM |
 | [`tts_demo.ipynb`](tts_demo.ipynb) | [![tts_demo](https://github.com/espnet/notebook/actions/workflows/tts_demo.yml/badge.svg)](https://github.com/espnet/notebook/actions/workflows/tts_demo.yml) | Type a sentence, hear it spoken — one English voice, then 128 of them |
 | [`enh_demo.ipynb`](enh_demo.ipynb) | [![enh_demo](https://github.com/espnet/notebook/actions/workflows/enh_demo.yml/badge.svg)](https://github.com/espnet/notebook/actions/workflows/enh_demo.yml) | Pull speech out of noise, and measure how much it helped |
 | [`spk_demo.ipynb`](spk_demo.ipynb) | [![spk_demo](https://github.com/espnet/notebook/actions/workflows/spk_demo.yml/badge.svg)](https://github.com/espnet/notebook/actions/workflows/spk_demo.yml) | Turn a voice into a vector, and score two recordings against each other |

Demos/s2t_pr_demo.ipynb

@@ -0,0 +1,236 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# Phones, and the two ways across to words\n",
+    "\n",
+    "[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/espnet/notebook/blob/master/Demos/s2t_pr_demo.ipynb) [![s2t_pr_demo](https://github.com/espnet/notebook/actions/workflows/s2t_pr_demo.yml/badge.svg)](https://github.com/espnet/notebook/actions/workflows/s2t_pr_demo.yml)\n",
+    "\n",
+    "[POWSM](https://arxiv.org/abs/2510.24992) is a phonetic foundation model:\n",
+    "it hears speech and answers in phones. POWSM-CTC is the encoder-only\n",
+    "variant, and it does four things with one recording —\n",
+    "\n",
+    "| task | you give | it answers |\n",
+    "| :-- | :-- | :-- |\n",
+    "| `<asr>` | the audio | the words |\n",
+    "| `<pr>` | the audio | the phones, in IPA |\n",
+    "| `<g2p>` | the audio **and the words** | the phones |\n",
+    "| `<p2g>` | the audio **and the phones** | the words |\n",
+    "\n",
+    "— which is what this notebook runs, one after the other, on the same six\n",
+    "seconds of speech. CPU is enough."
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Install"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "%pip install -q \"espnet==202610.post2\" espnet_model_zoo librosa"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## A recording"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import librosa\n",
+    "from IPython.display import Audio, display\n",
+    "\n",
+    "!wget -q -O sample.wav https://github.com/espnet/espnet/raw/master/test_utils/ctc_align_test.wav\n",
+    "speech, rate = librosa.load(\"sample.wav\", sr=16000)\n",
+    "display(Audio(speech, rate=rate))"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## The model\n",
+    "\n",
+    "There are two POWSMs, and this notebook runs on either: change `MODEL` below\n",
+    "and nothing else. `powsm_ctc` reads its window in one pass and is what\n",
+    "`espnet phonemize` loads; `powsm` is the encoder-decoder, which searches and\n",
+    "is about four times slower on a CPU. Both write the same phone set — a\n",
+    "diphthong is two symbols in either, by design — so what differs between them\n",
+    "is which phones they choose, not what they can say.\n",
+    "\n",
+    "`decode_window` is what makes the swap work: a CTC-only checkpoint is read\n",
+    "off its CTC head, one with a decoder is searched, and the caller does not\n",
+    "have to know which it is holding.\n",
+    "\n",
+    "POWSM reads a 20-second window, so a shorter clip is padded to it. The\n",
+    "checkpoint also carries its own spelling of \"work the language out yourself\"\n",
+    "- `<unk>` here, where OWSM writes `<nolang>` - so ask the model rather than\n",
+    "typing a symbol."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import librosa\n",
+    "from espnet2.bin.s2t_inference import Speech2Text\n",
+    "\n",
+    "MODEL = \"espnet/powsm_ctc\"  # or \"espnet/powsm\", the encoder-decoder\n",
+    "s2t = Speech2Text.from_pretrained(MODEL, device=\"cpu\")\n",
+    "\n",
+    "window = s2t.preprocessor_conf[\"speech_length\"]\n",
+    "nolang = s2t.no_language()\n",
+    "padded = librosa.util.fix_length(speech, size=rate * window)\n",
+    "print(f\"{window} s window, language symbol {nolang}, CTC-only: {s2t.ctc_only}\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## `<pr>`: the phones\n",
+    "\n",
+    "The task it is built for. POWSM writes each phone between slashes, so that\n",
+    "a phone spelled like a BPE token is still one token; spaced out is easier\n",
+    "to read and is what anything counting or aligning them wants."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import re\n",
+    "\n",
+    "\n",
+    "def phones(decoded):\n",
+    "    \"\"\"The phones out of POWSM's /p//h//o/ spelling.\"\"\"\n",
+    "    return \" \".join(re.findall(r\"/([^/]+)/\", decoded))\n",
+    "\n",
+    "\n",
+    "def run(task, text_prev=\"<na>\"):\n",
+    "    \"\"\"One window, decoded the way this checkpoint has to be.\"\"\"\n",
+    "    decoded = s2t.decode_window(padded, nolang, task, text_prev)\n",
+    "    return decoded.split(\">\")[-1].strip() if \"<\" in decoded else decoded\n",
+    "\n",
+    "\n",
+    "said_in_phones = phones(run(\"<pr>\"))\n",
+    "print(said_in_phones)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## `<asr>`: the words, for contrast\n",
+    "\n",
+    "A phonetic model will transcribe. On English — where a model trained for\n",
+    "text has seen far more — it is the weaker choice, and what comes back below\n",
+    "shows it. On a language that general speech corpora barely cover, the\n",
+    "comparison is a different one: POWSM is competitive there, because it was\n",
+    "trained on a phone-annotated collection rather than on whatever happens to\n",
+    "be plentiful."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "print(run(\"<asr>\"))"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## `<g2p>`: phones for words you give\n",
+    "\n",
+    "Now the audio is not the only input. The words you hand it ground the\n",
+    "answer — the phones come out for *those* words — and it is still listening\n",
+    "to the audio, which is where the pronunciation comes from."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "said = \"the sale of the hotels is part of holiday's strategy\"\n",
+    "\n",
+    "print(phones(run(\"<g2p>\", text_prev=said)))"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## `<p2g>`: words for phones you give\n",
+    "\n",
+    "The other direction. The prompt is phones, in the spelling the model was\n",
+    "trained on — each between slashes."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "as_prompt = \"/\" + \"//\".join(said_in_phones.split()) + \"/\"\n",
+    "print(as_prompt[:70], \"...\")\n",
+    "\n",
+    "print(run(\"<p2g>\", text_prev=as_prompt))"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Where next\n",
+    "\n",
+    "- **From the terminal**: `espnet phonemize sample.wav` is `<pr>`, in one line\n",
+    "- **In the browser**: the [powsm-ctc Space](https://huggingface.co/spaces/espnet/powsm-ctc)\n",
+    "  offers all four tasks, and `espnet demo --model espnet/powsm_ctc` is the same page\n",
+    "- **From an assistant**: the MCP server's `phonemize` tool\n",
+    "- **Alignment**, with the same kind of model: [`s2t_align_demo.ipynb`](s2t_align_demo.ipynb)\n",
+    "- **The encoder-decoder POWSM**, [`espnet/powsm`](https://huggingface.co/espnet/powsm),\n",
+    "  answers the same four tasks with a search rather than a CTC pass"
+   ]
+  }
+ ],
+ "metadata": {
+  "colab": {
+   "provenance": []
+  },
+  "kernelspec": {
+   "display_name": "Python 3",
+   "name": "python3"
+  },
+  "language_info": {
+   "name": "python"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 0
+}

README.md

@@ -18,6 +18,7 @@ One per task, flat in [`Demos/`](Demos), each short enough to read in a sitting.
 | [`asr_streaming_demo.ipynb`](Demos/asr_streaming_demo.ipynb) | [![asr_streaming_demo](https://github.com/espnet/notebook/actions/workflows/asr_streaming_demo.yml/badge.svg)](https://github.com/espnet/notebook/actions/workflows/asr_streaming_demo.yml) | Watch the words appear while the audio is still arriving |
 | [`st_demo.ipynb`](Demos/st_demo.ipynb) | [![st_demo](https://github.com/espnet/notebook/actions/workflows/st_demo.yml/badge.svg)](https://github.com/espnet/notebook/actions/workflows/st_demo.yml) | Translate English speech into German, French and Chinese — the same model |
 | [`s2t_align_demo.ipynb`](Demos/s2t_align_demo.ipynb) | [![s2t_align_demo](https://github.com/espnet/notebook/actions/workflows/s2t_align_demo.yml/badge.svg)](https://github.com/espnet/notebook/actions/workflows/s2t_align_demo.yml) | Line text up with the audio it was said in, and score how well they agree |
+| [`s2t_pr_demo.ipynb`](Demos/s2t_pr_demo.ipynb) | [![s2t_pr_demo](https://github.com/espnet/notebook/actions/workflows/s2t_pr_demo.yml/badge.svg)](https://github.com/espnet/notebook/actions/workflows/s2t_pr_demo.yml) | Hear the phones, and cross between phones and words with POWSM |
 | [`tts_demo.ipynb`](Demos/tts_demo.ipynb) | [![tts_demo](https://github.com/espnet/notebook/actions/workflows/tts_demo.yml/badge.svg)](https://github.com/espnet/notebook/actions/workflows/tts_demo.yml) | Type a sentence, hear it spoken — one English voice, then 128 of them |
 | [`enh_demo.ipynb`](Demos/enh_demo.ipynb) | [![enh_demo](https://github.com/espnet/notebook/actions/workflows/enh_demo.yml/badge.svg)](https://github.com/espnet/notebook/actions/workflows/enh_demo.yml) | Pull speech out of noise, and measure how much it helped |
 | [`spk_demo.ipynb`](Demos/spk_demo.ipynb) | [![spk_demo](https://github.com/espnet/notebook/actions/workflows/spk_demo.yml/badge.svg)](https://github.com/espnet/notebook/actions/workflows/spk_demo.yml) | Turn a voice into a vector, and score two recordings against each other |