syvai/subtitle-speak

Subtitle Speak

52,143 clips / 167 hours of expressive synthetic Danish speech, generated with ElevenLabs Eleven v4 from Danish text (mostly film and TV subtitles), with inline audio tags for emotion, delivery, non-verbal sounds, background environments and recording channels.

Two speakers, one male and one female: ElevenLabs Voice Library voices Søren and Freja. 32,880 clips are monologues and 19,263 are two-speaker dialogues between them, for 65,583 Søren turns and 65,377 Freja turns.

How it was made

1. Text. Units were cut from Danish Dynaword subsets and filtered to Danish only. A unit is 2–6 consecutive lines for a monologue, or 3–8 lines for a dialogue, with speakers alternating. Sources by clip: opensubtitles 32,218, hest 5,739, tv2r 4,791, danske-taler 3,318, wikipedia 2,943, wiki-comments 2,017, spont 723, hvadvilduhelst 394.
2. Tags. Every clip carries at least 2 tags, including at least 1 emotion, delivery or non-verbal tag. Tags are drawn evenly from fixed lists, so rare tags get as much coverage as common ones, and are inserted at sentence starts and boundaries. 33,854 clips (65%) open with a background environment and 6,812 (13%) with a recording-channel tag. In dialogues, about 12% of turns cut the previous speaker off with — and an interruption tag.
3. Audio. Each clip is one request to ElevenLabs Text to Dialogue with model_id=eleven_v4, stability 0.5 and similarity 0.75, saved as 44.1 kHz, 192 kbps MP3. A dialogue clip is a single audio file containing all of its turns.

Accent and character-voice tags, and one-off sound effects such as [door slams] or [gunshot], were deliberately left out.

Fields

field	description
id	clip id
audio	MP3 audio
kind	monologue or dialogue
speakers	speakers in the clip
text	exactly the tagged text sent to the model; dialogue turns are joined as Speaker: text lines
text_plain	the same text with all [tags] removed
turns	per turn: speaker, tagged text, text_plain
tags	every tag in the clip as {type, tag}; types are emotion, delivery, nonverbal, pause, turn, environment and channel
environment, channel	the scene-level tag, or null
source	Danish Dynaword subset the text came from

Tag inventory

Counts of tag occurrences: emotion 76,956, nonverbal 62,273, delivery 41,061, environment 33,854, pause 11,514, channel 6,812, turn 6,555.

• Emotion (60): [amused], [angry], [annoyed], [anxious], [appalled], [awe], [bitter], [bored], [calm], [cheerful], [confused], [curious], [defensive], [delighted], [determined], [disappointed], [disgusted], [dismissive], [elated], [embarrassed], [enthusiastic], [excited], [frustrated], [furious], [guilty], [happy], [heartbroken], [hopeful], [impatient], [irritated], [jealous], [melancholic], [mischievously], [nervous], [nostalgic], [panicking], [playful], [pleading], [proud], [reassuring], [relieved], [resigned], [sad], [sarcastic], [scared], [serious], [sheepishly], [shy], [skeptical], [sorrowful], [stunned], [surprised], [sympathetic], [teasing], [tender], [terrified], [tired], [uncertain], [warmly], [worried]
• Delivery (31): [breathy], [crying while speaking], [deadpan], [dramatically], [drawn out], [firmly], [hesitant], [laughing while speaking], [loudly], [low, threatening], [matter-of-fact], [monotone], [mumbling], [out of breath], [quietly], [rushed], [shouting], [singing], [sleepy], [softly], [speaking slowly], [speaking through a yawn], [speaking very fast], [stuttering], [trailing off], [under her breath], [under his breath], [whispers], [with a smile in the voice], [with a trembling voice], [yelling]
• Non-verbal (40): [breathing heavily], [chuckles], [clears throat], [coughs], [crying], [exhales], [exhales sharply], [frustrated sigh], [gasps], [giggles], [groans], [grunts], [gulps], [happy gasp], [hiccups], [hums], [inhales deeply], [laughs], [laughs harder], [laughs hysterically], [muttering], [nervous laugh], [panting], [scoffs], [sighs], [smacks lips], [sneezes], [snickers], [sniffs], [snorts], [sobs], [soft chuckle], [starts laughing], [stifling laughter], [swallows], [tsks], [wheezing], [whimpers], [whistles], [yawns]
• Pause: [long pause], [pause], [short pause]
• Turn-taking: [cutting in], [interrupting], [jumping in], [overlapping]
• Environment (38): [air conditioning hum], [airplane cabin hum], [beach with ocean waves], [busy café, chatter and clinking cups], [busy market], [busy restaurant], [city square with crowd], [concert hall before a show], [construction site, drilling], [crowded bar with music], [crowded party], [fan running], [farm with animals], [fireplace crackling], [forest], [gym with music], [harbour with seagulls], [heavy rain], [heavy street traffic], [hospital corridor], [inside a bus], [inside a moving car], [inside a moving train], [kitchen, cooking sounds], [light rain], [living room with TV on], [night with crickets], [open-plan office], [park with birds singing], [playground with children], [quiet countryside], [school hallway], [snowstorm], [stadium crowd cheering], [strong wind outdoors], [supermarket], [thunderstorm], [train station with announcements]. Phrased as [X in the background], [X, ambient noise throughout] or [background: X].
• Channel (10): [far away from the microphone], [in a large echoing hall], [in a large empty church], [in a small car], [in a small tiled bathroom], [muffled, through a wall], [on a phone call, telephone audio quality], [on a video call, compressed audio], [over a crackly walkie-talkie], [speaking through a megaphone]

Limitations

• Tags are placed at random. They don't follow the meaning of the text, so cheerful lines can carry [sobs] and stacked tags can contradict each other.
• No quality check. Nothing was transcribed, scored or listened through, so a tag may not be audible in its clip, and the audio may occasionally drift from the text.
• Subtitle text quality. Some subtitle lines read as weak translations.
• No timestamps. Dialogue audio has no turn or word timings.
• Excluded clips. 165 planned clips are missing: 34 were blocked by ElevenLabs' content filter and the rest failed when credits ran out. Another 723 clips were left out because a bug sent Freja's [whispers] to the model as [wherpers], so the female whisper tag is underrepresented.

License

• Text comes from Danish Dynaword (CC0-1.0); see the individual subsets for their provenance.
• Audio was generated with ElevenLabs Eleven v4 and is subject to the ElevenLabs Terms of Service. Søren and Freja are Voice Library voices owned by their voice creators. Check both before reusing the audio, especially for training speech models.

Created by syv.ai.