Melati

Multilingual Expressive speech, Labelled for Affect, Tags and Interjections: 591 hours of Malaysian podcast speech in 355,477 segments, each with its transcript, word times, speaker, quality scores and tags for emotion, non-verbal sounds (breaths, laughs, coughs, sniffs, throat clearing), whispering and fillers, written in Fish Audio Drama 3's tag syntax:

<happy>Tapi bukan jari aku, tu jaring apa, kita pun tak [breathes] tahu.</happy> <fearful>Tak tahu, tapi air keluar daripada situ.</fearful>

(segment r416c336e96a5_0164.) Built by Melati, an in-the-wild speech-data pipeline in the line of Emilia-Pipe, from the recordings of malaysia-ai/malaysian-podcast-youtube, the podcasts behind Malaysian-Emilia. No score is filtered on: every measurement is a column, and a subset is a query (below). The only rows left out are those whose transcript cannot be what was said (8,010 segments, 4.6 h: below the languages).

A partial release, growing. This holds 6,912 of the source's 19,092 recordings, 108 of its 299 shards; the rest is being processed and is added here as it finishes. Every number on this card is over what is here.

source recordings	6,912 of 19,092 (833 h; 50 of the source's files are empty)
segments	355,477
hours	591.4
speaker ids	14,686, per recording, or per ten-minute window in the 219 recordings (182 h) diarized in windows
audio	24 kHz mono FLAC, inline in the parquet shards

language	hours	segments
Malay	565.2	326,070
English	26.3	29,407

Left out, as a transcript that cannot be what was said:

why	segments	hours
Tamil or Mandarin the ASR model did not hear itself: when it detects a language outside Malay, English, Mandarin and Tamil, the segment is decoded in each of those and the best-scoring kept, and on podcasts that choice lands on Tamil for Arabic chant, songs and stray Malay	6,939	4.0
Tamil or Mandarin heard in a recording mostly in another language: songs, chants and a stray phrase	972	29 min
no text at all: the aligner placed none of the words	37	5 min
text whose letters are mostly in a script its language is not written in	62	4 min

Use

from datasets import load_dataset

ds = load_dataset("Scicom-intl/Melati", split="train", streaming=True)
for row in ds:
    audio = row["audio"]["array"]           # 24 kHz
    print(row["tagged_text"], row["language"], row["mos"])
    break

Subsets are filters on the columns. The core of Emilia-Pipe's filter (3 to 30 s, DNSMOS at least 3) keeps 86,074 segments, 180.6 h:

emilia_like = ds.filter(lambda r: 3 <= r["duration"] <= 30 and (r["dnsmos"] or 0) >= 3)
laughs = ds.filter(lambda r: r["n_laughs"] > 0)

Tags

Per hour of speech: non-verbal sounds (blue), fillers (teal), emotion spans (amber).

Every tag is a rendering of time-stamped events kept in annotation, so a different target format needs no re-annotation. A tag class appears only where its precision was measured on labelled speech no model trained on, at a threshold chosen on other labelled speech; every threshold is in export_config.json.

• Emotion spans (<happy>...</happy>; neutral is untagged): the average of seven models, Melati's heads on five frozen encoders (one released) and MERaLiON-SER and emotion2vec+, over phrases between pauses. Thresholds (happy 0.51, sad 0.64, angry 0.48, surprised 0.3, fearful 0.44) give 0.75 precision at a podcast's assumed label rate on held-out English and Polish speech (nEMO, NonverbalTTS); chosen on one half and measured on the other they hold 0.74 to 0.82. Disgust is never tagged.
• Non-verbal point tags. Breaths ([breathes], 52.1 an hour) come from Scicom-intl/Melati-NonVerbal-Whisper-Large-V3's gap head in the gaps between aligned words (threshold 0.7); laughs ([laughs], 29.5 an hour, 5,007 of them over speech) and coughs ([coughs], 0.20 an hour) from its frame head, every 0.1 s, over speech too (laugh 0.97, cough 0.999), and from AudioSet's taggers (laugh 0.3, cough 0.5); sniffs ([sniffs], 0.90 an hour, threshold 0.15) and throat clearing ([clears throat], 0.09 an hour, 0.2) from AudioSet's taggers alone: the average of PretrainedSED's BEATs, ATST-F and M2D, fine-tuned on AudioSet's time-stamped labels, every 0.1 s. On PodcastFillers' test episodes through this pipeline the tags find 70% of breaths and 71% of laughs, 83% and 84% precise where labelled. AudioSet's thresholds were each chosen on one half of NonverbalTTS's clips (interviews and expressive speech no model here trained on) and measured on the other: sniffs 85 to 88% precise, 68 to 71% found; throat clearing 81 to 83% precise, 34 to 47% found; coughs 80 to 97% precise, 17 to 28% found; laughs 84 to 87% precise, 45 to 46% found.
• Delivery: whispering (<whisper>...</whisper>, 0.02 spans an hour) over the words AudioSet's taggers hear whispered (threshold 0.3): on Expresso they cover 99% of the whispered utterances and 0% of the 10,097 in its other styles.
• Fillers are words (um, erm, emm, 201 an hour), written by the ASR model or recovered from a voiced gap transcribed alone; tagged_text_hesitates writes them as [hesitates].

Quality

Each segment is exported from the recording itself, or, where the separator heard music, from the separated version (speech and vocal sounds kept, music removed: a speech-only stem would delete the laughs and breaths the tags describe), whichever SOTA-MOS scores higher. 56,449 segments are the separated version. SOTA-MOS median 2.943, DNSMOS median 2.898.

Columns

column	what
audio	the segment, 24 kHz mono FLAC (16-bit), cut from its first sound to its last plus 0.15 s
id, recording, speaker	segment id; source recording; speaker within that recording (diarized per recording, so one person in two recordings has two ids; a recording over 45 minutes (80 for those finished before 2026-10-26), or whose single pass failed, is diarized in ten-minute windows and its ids, <recording>_w<k>S<j>, are per window)
start, end, duration	where the segment sits in the source recording, in seconds
language	the language the transcript was decoded in (ms, en, ta, zh; Tamil and Mandarin only where the ASR model heard them itself, in a recording mostly in that language)
text	the transcript as words: fillers kept as spoken (um, erm, emm), no tags
tagged_text	the transcript with tags in Fish Audio Drama 3's syntax: emotion scopes and non-verbal point tags
tagged_text_hesitates	the same with each filler written as [hesitates]
annotation	JSON: every word with its time and confidence, every non-verbal event with its time, probability and source, every emotion span with its words and probability. The tagged text is a rendering of this
asr_text	the transcript as the ASR model wrote it, before fillers found in the gaps were added
emotion_segment, emotion_segment_p	the most probable emotion of the whole segment (the seven-model average) and its probability, whatever the threshold
n_events, n_breaths, n_laughs, n_coughs, n_sniffs, n_throat_clears, n_fillers	counts, for filtering
whispered	the segment holds a whispered span (<whisper>...</whisper> in the tagged text)
n_dropped_words	voiced gaps between aligned words that, transcribed alone, came back as a word the transcript does not have: mostly backchannels (ha, ya, okay) and short function words written for a breath or a click, so a count to filter on, not words missing from the text
squeezed_words	words the aligner fit into under 25 ms (one frame): text the audio does not hold, usually words the ASR model wrote past the end of what was cut. squeezed_words < 2 is the stricter set
audio_version	mix (the recording itself) or vocal (music removed by DBS-TIGER, speech and vocal sounds kept), whichever SOTA-MOS prefers where there was music
mos, mos_vocal, mos_mix	SOTA-MOS of the exported audio, of the separated version and of the mix
dnsmos	DNSMOS P.835 overall score of the exported audio (Emilia's filter)
lufs, peak_db, background_db	loudness, peak, and music level against speech in the segment
align_score, edge_clip_start, edge_clip_end	forced-alignment confidence; a word starting or ending within 30 ms of the cut
forced_cut, overlap_s_removed	a segment cut inside speech because it ran past 30 s; seconds of overlapping speech removed
asr_avg_logprob, asr_fallback_t, asr_looped	decoder confidence; the temperature a looping transcript was re-decoded at (0 when greedy was kept); a loop survived
prev, next	the neighbouring segments of the same recording, for context
source, source_title, youtube_id	where the audio comes from

How it was built

stage	model
decode, one gain per recording; music separation only where PANNs hears background	PANNs Cnn14, DBS-TIGER
who spoke when	Who-Spoke-When-Over-the-Phone
single-speaker segments, cut only in pauses	Silero VAD
transcription	Scicom-intl/Melati-Whisper-Large-V3
word times	Where-Words-Begin-and-End
quality	SOTA-MOS (24 kHz), DNSMOS P.835
sniffs, throat clearing, whispering, more laughs and coughs	PretrainedSED's BEATs, ATST-F and M2D on AudioSet's strong labels, averaged
emotion, non-verbal, fillers	above

The code, the evaluation of every model and the comparison with Emilia-Pipe on the same recordings are in Melati on GitHub.

Limitations

• The transcripts, speakers, tags and scores are model output. Every tag's precision was measured on English (and Polish, for emotion) speech; on Malay speech it is not yet measured against human labels.
• Coughs, sniffs, throat clearing and whispering are rare in podcasts and rarer as tags: each is tagged only where its precision held, so most are missed. Sighs, gasps, crying, humming, singing and shouting are not tagged yet: AudioSet's taggers were imprecise on them, never heard them in podcasts, or (sighs) there is no labelled conversational speech to measure them on.
• Speaker ids hold within one recording only, and within one ten-minute window in a recording diarized in windows.

Licence

CC BY-NC 4.0, as the source. The audio belongs to its owners; it is shared for non-commercial research, under the source's own notice:

All the videos, songs, images, and graphics used in the video belong to their respective owners and I does not claim any right over them. Copyright Disclaimer under section 107 of the Copyright Act of 1976, allowance is made for "fair use" for purposes such as criticism, comment, news reporting, teaching, scholarship, education and research.

Citation

@misc{zolkepli2026melati,
  title  = {Melati: Multilingual Expressive speech, Labelled for Affect, Tags and Interjections},
  author = {Husein Zolkepli},
  year   = {2026},
  note   = {Scicom (MSC) Berhad}
}