YODAS-2-MT / dashboard.html
Lakoc's picture
Recipe, statistics and sample list for the YODAS2 multi-talker pseudo-labels (CC-BY-4.0)
ef65191 verified
Raw History Blame Contribute Delete
17.1 kB
<!doctype html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1, viewport-fit=cover">
<title>YODAS2 Pseudo-Labels</title>
<link rel="preconnect" href="https://fonts.googleapis.com">
<link rel="stylesheet" href="https://fonts.googleapis.com/css2?family=Archivo:wght@500;700;800&family=Source+Sans+3:wght@400;600&family=JetBrains+Mono:wght@400;600&display=swap">
<style>
/* Layout: one reading column; a figure strip, then chart panels in a two-column grid that stacks on phones. */
:root {
--bg: #f3f5f6; --surface: #ffffff; --fg: #17242b; --muted: #5b6b73; --line: #d7dee2;
--accent: #1d6b75; --accent-soft: #cfe4e6; --warn: #a65a12; --warn-soft: #f4e2cf;
--display: "Archivo", "Helvetica Neue", Arial, sans-serif;
--body: "Source Sans 3", "Segoe UI", system-ui, sans-serif;
--mono: "JetBrains Mono", ui-monospace, "SFMono-Regular", Menlo, monospace;
}
@media (prefers-color-scheme: dark) { :root:not([data-theme="light"]) {
--bg: #11181c; --surface: #172127; --fg: #e3ebee; --muted: #93a4ac; --line: #2a3840;
--accent: #5fb6c0; --accent-soft: #1f3a3f; --warn: #e3a25e; --warn-soft: #3a2a1a; color-scheme: dark } }
:root[data-theme="dark"] {
--bg: #11181c; --surface: #172127; --fg: #e3ebee; --muted: #93a4ac; --line: #2a3840;
--accent: #5fb6c0; --accent-soft: #1f3a3f; --warn: #e3a25e; --warn-soft: #3a2a1a; color-scheme: dark }
body { background: var(--bg); color: var(--fg); font: 16px/1.55 var(--body); }
.wrap { max-width: 1040px; margin: 0 auto; padding-block: 40px 64px;
padding-left: max(16px, env(safe-area-inset-left, 0px)); padding-right: max(16px, env(safe-area-inset-right, 0px));
display: grid; gap: 36px; }
h1, h2 { font-family: var(--display); text-wrap: balance; margin: 0; }
h1 { font-size: clamp(2rem, 5vw, 2.9rem); font-weight: 800; letter-spacing: -0.01em; line-height: 1.1; }
h2 { font-size: 1.25rem; font-weight: 700; }
p { margin: 0; max-width: 68ch; }
.eyebrow { font: 600 0.75rem/1 var(--mono); letter-spacing: 0.08em; text-transform: uppercase; color: var(--accent); }
header { display: grid; gap: 14px; }
.lede { color: var(--muted); font-size: 1.08rem; }
.notice { border: 1px solid var(--warn); background: var(--warn-soft); border-radius: 6px; padding: 14px 16px;
display: grid; gap: 6px; max-width: 78ch; }
.notice strong { color: var(--warn); font-family: var(--display); letter-spacing: 0.01em; }
.figures { display: grid; grid-template-columns: repeat(auto-fit, minmax(150px, 1fr)); gap: 1px;
background: var(--line); border: 1px solid var(--line); border-radius: 6px; overflow: hidden; }
.fig { background: var(--surface); padding: 16px; display: grid; gap: 4px; align-content: start; }
.fig b { font: 600 1.6rem/1.1 var(--mono); font-variant-numeric: tabular-nums; }
.fig span { color: var(--muted); font-size: 0.9rem; }
section { display: grid; gap: 14px; }
.funnel { display: grid; gap: 8px; }
.step { display: grid; grid-template-columns: minmax(0, 15rem) minmax(0, 1fr); gap: 12px; align-items: center; }
.step .what { font-size: 0.95rem; }
.step .what small { display: block; color: var(--muted); font-size: 0.82rem; }
.bar { height: 28px; background: var(--accent-soft); border-radius: 3px; position: relative; min-width: 2px; }
.bar.on { background: var(--accent); }
.bar em { position: absolute; left: 8px; top: 50%; transform: translateY(-50%); font: 600 0.82rem/1 var(--mono);
font-style: normal; color: var(--fg); white-space: nowrap; font-variant-numeric: tabular-nums; }
.bar.on em { color: var(--bg); }
.grid { display: grid; grid-template-columns: repeat(2, minmax(0, 1fr)); gap: 20px; }
.panel { background: var(--surface); border: 1px solid var(--line); border-radius: 6px; padding: 18px; display: grid;
gap: 10px; align-content: start; min-width: 0; }
.panel.wide { grid-column: 1 / -1; }
.panel p.note { color: var(--muted); font-size: 0.88rem; }
svg { display: block; width: 100%; height: auto; overflow: visible; }
svg text { fill: var(--muted); font: 11px var(--mono); }
svg .val { fill: var(--fg); font-weight: 600; }
svg .mark { fill: var(--accent); }
svg .mark.dim { fill: var(--accent-soft); }
svg .axis { stroke: var(--line); stroke-width: 1; }
table { border-collapse: collapse; width: 100%; font-size: 0.92rem; }
th, td { text-align: left; padding: 7px 10px; border-bottom: 1px solid var(--line); }
th { font: 600 0.75rem/1.2 var(--mono); text-transform: uppercase; letter-spacing: 0.06em; color: var(--muted); }
td.n { text-align: right; font-family: var(--mono); font-variant-numeric: tabular-nums; }
.tablebox { overflow-x: auto; }
footer { color: var(--muted); font-size: 0.88rem; display: grid; gap: 6px; border-top: 1px solid var(--line); padding-top: 18px; }
a { color: var(--accent); }
code { font-family: var(--mono); font-size: 0.88em; }
@media (max-width: 720px) {
.grid { grid-template-columns: minmax(0, 1fr); }
.step { grid-template-columns: minmax(0, 1fr); gap: 4px; }
}
</style>
<style>body{margin:0}</style>
</head>
<body>
<div class="wrap">
<header>
<div class="eyebrow">Pandemonium · training data, stage 2</div>
<h1>YODAS2 Pseudo-Labels</h1>
<p class="lede">Multi-speaker YouTube conversations from <a href="https://huggingface.co/datasets/espnet/yodas2">YODAS2</a>,
cut into chunks of up to 10 minutes and transcribed speaker by speaker with
<a href="https://huggingface.co/BUT-FIT/DiCoW_v3_3">DiCoW v3.3</a> conditioned on an automatic diarization.
These are the statistics of the 4.6k hours that Pandemonium was trained on.</p>
<div class="notice" role="note">
<strong>Not redistributed</strong>
<p>Because of possible licensing issues with the source recordings, neither the audio nor the labels are
shared. The corpus is released as the scripts that mine it, plus a list of 100 sample recordings, so that it
can be rebuilt. The public recipe uses Nemotron-3-Diarization and DiCoW v3.3 and gives a comparable corpus,
not an identical one.</p>
</div>
</header>
<div class="figures" id="figures"></div>
<section>
<h2>From YODAS2 to the training set</h2>
<div class="funnel" id="funnel"></div>
<p class="lede" style="font-size:0.95rem">Recordings were drawn at random, stratified by language, from those
with at least 2 speakers, 5 s of overlapped speech and 15 s from the quietest speaker. Chunks with more than 10
speakers were dropped; the labels identify at most 8. The last step keeps the languages Parakeet-v3 supports,
the only filter applied to Pandemonium's training data.</p>
</section>
<div class="grid">
<div class="panel wide">
<h2>Hours per language</h2>
<div id="lang"></div>
<p class="note" id="langnote"></p>
</div>
<div class="panel">
<h2>Speakers per chunk</h2>
<div id="spk"></div>
<p class="note">Hours of audio by the number of speakers the diarization found in a chunk.</p>
</div>
<div class="panel">
<h2>Overlapped speech</h2>
<div id="ovl"></div>
<p class="note" id="ovlnote"></p>
</div>
<div class="panel">
<h2>Chunk length</h2>
<div id="dur"></div>
<p class="note">Number of chunks by length in minutes. Recordings are cut at silences near every 10 minutes,
so most chunks are full length and the rest are recording tails.</p>
</div>
<div class="panel">
<h2>Share of the main speaker</h2>
<div id="dom"></div>
<p class="note" id="domnote"></p>
</div>
</div>
<section>
<h2>What the optional filters would remove</h2>
<p>The paper's DiCoW study (Table 2) also tested three label filters. They are in the recipe as optional steps;
Pandemonium was trained without them. Run on this corpus, the two rule-based ones mark:</p>
<div class="tablebox"><table id="filters"></table></div>
<p class="lede" style="font-size:0.95rem">A flagged chunk keeps its unflagged windows of at least 2 minutes; an
unsalvageable one has none. The third filter, an LLM language check, needs an API and is not counted here.</p>
</section>
<footer>
<p>Rebuild it: the recipe, the sample list and the scripts are in <code>data/yodas_pseudo/</code> of the
<a href="https://github.com/BUTSpeechFIT/Pandemonium">Pandemonium repository</a>. Method: section 3.1.1 of the
<a href="https://huggingface.co/BUT-FIT/Pandemonium/blob/main/Pandemonium_ICASSP2027.pdf">paper</a>.</p>
<p>Hours are chunk durations; speech, overlap and speaker counts come from the diarization the labels were
conditioned on. Language is the YODAS2 shard's.</p>
</footer>
</div>
<script>
const D = {"kpi": {"chunks": 32865, "recordings": 11862, "hours": 4627.2, "speech_hours": 3867.5, "languages": 23, "words": 45939448, "mean_speakers": 3.19, "median_chunk_min": 10.0, "overlap_hours": 148.2}, "full": {"chunks": 38399, "recordings": 14682, "hours": 5362.6, "languages": 79}, "lang": [["en", 2114.8, 14875], ["es", 842.7, 6073], ["ru", 595.9, 4235], ["pt", 337.3, 2400], ["fr", 283.6, 2032], ["de", 181.3, 1270], ["it", 171.9, 1250], ["nl", 56.2, 420], ["pl", 15.0, 106], ["uk", 10.5, 77], ["cs", 5.0, 33], ["hu", 3.0, 22], ["bg", 2.2, 15], ["fi", 2.1, 15], ["sv", 1.9, 14], ["el", 1.6, 12], ["da", 0.5, 4], ["sk", 0.5, 4], ["sl", 0.4, 3], ["ro", 0.3, 2], ["lv", 0.2, 1], ["lt", 0.1, 1], ["hr", 0.1, 1]], "captions": {"manual": 589.6, "auto": 4037.6}, "speakers": {"1": 53.6, "2": 1957.2, "3": 1093.6, "4": 692.5, "5": 379.5, "6": 218.9, "7": 135.3, "8": 96.6}, "speakers_sup": {"1": 457, "2": 14353, "3": 7805, "4": 4841, "5": 2536, "6": 1444, "7": 845, "8": 584}, "overlap": [["[0.0, 0.01)", 1928.1], ["[0.01, 0.02)", 722.0], ["[0.02, 0.04)", 743.3], ["[0.04, 0.06)", 422.8], ["[0.06, 0.08)", 270.3], ["[0.08, 0.1)", 164.9], ["[0.1, 0.15)", 202.4], ["[0.15, 0.2)", 78.1], ["[0.2, 1.01)", 95.2]], "overlap_stats": {"median": 0.016, "mean": 0.037, "zero": 8.5, "median_sec": 5.8}, "duration": [["0-1", 668], ["1-2", 891], ["2-3", 1066], ["3-4", 1190], ["4-5", 1244], ["5-6", 1165], ["6-7", 1199], ["7-8", 1149], ["8-9", 1119], ["9-10", 17539]], "dominant": {"median": 0.642}, "dominant_hist": [660, 2332, 3589, 7264, 5969, 5275, 4203, 3573], "wps": 3.3, "hallucination_verdict": [["clean", 26994, 3776.0], ["flagged", 5625, 829.2], ["would_quarantine_unsalvageable", 246, 22.1]], "leakage_verdict": [["clean", 30909, 4339.5], ["flagged", 1885, 284.0], ["would_quarantine_leakage", 71, 3.7]], "both_clean": [25688, 3584.6]};
const fmt = (x, d = 0) => Number(x).toLocaleString("en-US", { minimumFractionDigits: d, maximumFractionDigits: d });
const NS = "http://www.w3.org/2000/svg";
const el = (tag, attrs, text) => { const e = document.createElementNS(NS, tag);
for (const k in attrs) e.setAttribute(k, attrs[k]); if (text != null) e.textContent = text; return e; };
const k = D.kpi;
document.getElementById("figures").innerHTML = [
[fmt(k.hours, 1) + " h", "audio"], [fmt(k.chunks), "chunks"], [fmt(k.recordings), "YouTube recordings"],
[k.languages, "languages"], [fmt(k.words / 1e6, 1) + " M", "words"], [k.mean_speakers, "speakers per chunk, mean"],
].map(([v, l]) => `<div class="fig"><b>${v}</b><span>${l}</span></div>`).join("");
const steps = [
["Recordings with diarization statistics", "YODAS2, all languages", null, "514,406 recordings"],
["Selected for labeling", "random, stratified by language", 5528.8, "15,086 recordings · 5,528.8 h"],
["Chunked and labeled", D.full.languages + " languages", D.full.hours, fmt(D.full.chunks) + " chunks · " + fmt(D.full.hours, 1) + " h"],
["In Parakeet-v3 languages", "Pandemonium's training set", k.hours, fmt(k.chunks) + " chunks · " + fmt(k.hours, 1) + " h"],
];
document.getElementById("funnel").innerHTML = steps.map(([w, s, h, lab], i) => {
const width = h == null ? 100 : 100 * h / 5528.8;
return `<div class="step"><div class="what">${w}<small>${s}</small></div>
<div class="bar${i === steps.length - 1 ? " on" : ""}" style="width:${width}%"><em>${lab}</em></div></div>`;
}).join("");
// Horizontal bars: label column, bar, value. rows = [[label, value, extra]]
function hbars(id, rows, unit, opts = {}) {
const rowH = 22, labW = opts.labW || 44, valW = 150, W = 640, H = rows.length * rowH + 4;
const max = Math.max(...rows.map(r => r[1]));
const svg = el("svg", { viewBox: `0 0 ${W} ${H}`, role: "img", "aria-label": opts.label || id });
rows.forEach(([lab, v, extra], i) => {
const y = i * rowH + 2, w = Math.max(1.5, (W - labW - valW) * v / max);
svg.append(el("text", { x: labW - 8, y: y + 14, "text-anchor": "end" }, lab));
svg.append(el("rect", { x: labW, y: y + 3, width: w, height: rowH - 8, rx: 2, class: "mark" }));
svg.append(el("text", { x: labW + w + 6, y: y + 14, class: "val" }, fmt(v, opts.digits ?? 1) + unit + (extra ? " " + extra : "")));
});
document.getElementById(id).append(svg);
}
// Vertical bars with category labels underneath.
function vbars(id, cats, vals, unit, opts = {}) {
const W = 420, H = 200, pad = { l: 8, r: 8, t: 18, b: 34 };
const max = Math.max(...vals), n = vals.length, slot = (W - pad.l - pad.r) / n, bw = slot * 0.72;
const svg = el("svg", { viewBox: `0 0 ${W} ${H}`, role: "img", "aria-label": opts.label || id });
svg.append(el("line", { x1: pad.l, x2: W - pad.r, y1: H - pad.b, y2: H - pad.b, class: "axis" }));
vals.forEach((v, i) => {
const h = (H - pad.t - pad.b) * v / max, x = pad.l + i * slot + (slot - bw) / 2, y = H - pad.b - h;
svg.append(el("rect", { x, y, width: bw, height: Math.max(h, 1), rx: 2, class: "mark" + (opts.dim && opts.dim(i) ? " dim" : "") }));
svg.append(el("text", { x: x + bw / 2, y: y - 5, "text-anchor": "middle", class: "val" }, fmt(v, opts.digits ?? 0)));
svg.append(el("text", { x: x + bw / 2, y: H - pad.b + 15, "text-anchor": "middle" }, cats[i]));
});
if (opts.xlabel) svg.append(el("text", { x: W / 2, y: H - 4, "text-anchor": "middle" }, opts.xlabel));
document.getElementById(id).append(svg);
}
const big = D.lang.filter(r => r[1] >= 1), small = D.lang.filter(r => r[1] < 1);
hbars("lang", big.map(([l, h, c]) => [l, h, "· " + fmt(c) + " chunks"]), " h", { label: "Hours per language" });
document.getElementById("langnote").textContent =
"Under 1 h each: " + small.map(([l, h]) => `${l} ${h} h`).join(", ") +
". ISO 639-1 codes. Of the total, " + fmt(D.captions.manual, 1) + " h come from YODAS2's manually captioned splits " +
"(shards such as en000) and " + fmt(D.captions.auto, 1) + " h from the automatic-caption splits (en100 and up); " +
"the captions themselves are not used.";
const sk = Object.keys(D.speakers);
vbars("spk", sk, sk.map(s => D.speakers[s]), " h", { xlabel: "speakers", dim: i => sk[i] === "1", label: "Hours by speakers per chunk" });
const ocats = ["<1", "1–2", "2–4", "4–6", "6–8", "8–10", "10–15", "15–20", "≥20"];
vbars("ovl", ocats, D.overlap.map(r => r[1]), " h", { xlabel: "overlapped share of speech, %", label: "Hours by overlap share" });
const o = D.overlap_stats;
document.getElementById("ovlnote").textContent = `Hours of audio by the share of speech where two or more people talk at
once. Median ${fmt(o.median * 100, 1)} % (${fmt(o.median_sec, 1)} s per chunk), mean ${fmt(o.mean * 100, 1)} %;
${fmt(k.overlap_hours, 1)} h of overlapped speech in total, and ${o.zero} % of chunks have none.`;
vbars("dur", D.duration.map(r => r[0]), D.duration.map(r => r[1]), "", { xlabel: "minutes", label: "Chunks by length" });
const dcats = ["20–30", "30–40", "40–50", "50–60", "60–70", "70–80", "80–90", "90–100"];
vbars("dom", dcats, D.dominant_hist, "", { xlabel: "main speaker's share of speech, %", label: "Chunks by main speaker share" });
document.getElementById("domnote").textContent = `Number of chunks by how much of the speech the most talkative
speaker has. Median ${fmt(D.dominant.median * 100, 0)} %, so most chunks are conversations rather than a monologue
with interjections.`;
const vmap = Object.fromEntries;
const H_ = vmap(D.hallucination_verdict.map(r => [r[0], r])), L_ = vmap(D.leakage_verdict.map(r => [r[0], r]));
const row = (name, rule, f, q) => `<tr><td>${name}</td><td>${rule}</td>
<td class="n">${fmt(f[1])}</td><td class="n">${fmt(f[2], 1)} h</td><td class="n">${fmt(q[1])}</td><td class="n">${fmt(q[2], 1)} h</td></tr>`;
document.getElementById("filters").innerHTML = `<thead><tr><th>filter</th><th>rule</th><th>flagged chunks</th><th>their audio</th>
<th>unsalvageable</th><th>their audio</th></tr></thead><tbody>
${row("Hallucinations", "boilerplate phrases; a 2–4-word phrase repeated ≥ 6× over ≥ 60 % of a segment; ≥ 5 identical segments in a row", H_.flagged, H_.would_quarantine_unsalvageable)}
${row("Leakage", "two speakers share ≥ 6 consecutive words within 1 s", L_.flagged, L_.would_quarantine_leakage)}
<tr><td colspan="2">Clean under both</td><td class="n">${fmt(D.both_clean[0])}</td><td class="n">${fmt(D.both_clean[1], 1)} h</td><td></td><td></td></tr></tbody>`;
</script>
</body>
</html>