belztjti/au

Bilingual Speech-Command Synthetic Dataset

Dataset summary

This dataset contains 847.251 hours of TTS audio (35.302 days) distributed across 1,015,296 audio samples. Each sample pairs embedded audio with its transcript, language, speaker identifier, and target string.

Scale and audited totals

Measure	Value
Total audio duration	3,050,103.745 seconds
Total audio duration	847.251 hours / 35.302 days
Audio samples	1,015,296
Mean duration per sample	3.004 seconds
Source texts	512
Voices	1,983
Original WAV storage	292.891 GB

Target representation

Each sample has one target string. A target consists of an exact command token and, for timed movement commands, an optional positive millisecond operand appended directly to the token.

Target	Meaning	Samples	Share of all samples
<|cmd_lower_compression|>	Lower the compression plate or paddle	126,912	12.5%
<|cmd_raise_compression|>	Raise the compression plate or paddle	103,116	10.2%
<|cmd_lower_swivel|>	Lower the swivel arm	118,980	11.7%
<|cmd_raise_swivel|>	Raise the swivel arm	126,912	12.5%
<|cmd_stop|>	Stop the current or named movement	111,048	10.9%
<|cmd_register_operator|>	Register the current voice as the operator	23,796	2.3%
<|cmd_none|>	No single supported target can be resolved	404,532	39.8%

The six command targets account for 610,764 samples (60.2%). <|cmd_none|> accounts for 404,532 samples (39.8%).

Command-target coverage

The command-target samples include terse commands, explicit and implicit movement durations, continuous actions, continuation and repetition language, global and direction-specific stopping, and corrections whose final intent is complete.

Command-target group	Samples	Share within command-target samples	Semantic forms
Movement	475,920	77.9%	210,198 brief, default, repeated, or continued movements mapped to 700 ms; 190,368 movements with an explicit positive duration; 75,354 continuous movements with no operand
<|cmd_stop|>	111,048	18.2%	27,762 generic or current-movement stops; 63,456 component-and-direction-specific stops; 19,830 global or multi-movement stops
<|cmd_register_operator|>	23,796	3.9%	15,864 direct registrations; 7,932 corrections whose final complete intent is registration

The 190,368 explicitly timed movement samples cover 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0, 1.2, 1.5, 2.0, 2.5, 10, and 30 seconds. Natural formulations include milliseconds, whole and fractional seconds, decimal phrasing, and longer upper-bound durations. Brief, default, repeated, or continued movements use a 700 ms operand when no numeric duration is spoken. Continuous movement targets omit the operand.

<|cmd_none|> semantic coverage

The 404,532 <|cmd_none|> samples are divided into 15 mutually exclusive primary semantic situations. The percentages below use all <|cmd_none|> samples as the denominator.

Primary semantic situation	Samples	Share	Detailed coverage
Negation	51,558	12.7%	23,796 prohibited movements; 23,796 requests not to stop; 3,966 requests not to register the voice
Multiple commands or alternatives	27,762	6.9%	15,864 paired supported movements; 3,966 movement alternatives; 3,966 registration-plus-movement combinations; 3,966 supported-plus-unsupported combinations
Conflict or reversal	15,864	3.9%	Stop-then-reverse sequences
Invalid duration	15,864	3.9%	3,966 zero-duration movements; 3,966 numbers without a time unit; 3,966 distance operands; 3,966 alternative durations
Delayed or periodic scheduling	7,932	2.0%	3,966 delayed requests and 3,966 periodically repeated requests
Questions	27,762	6.9%	11,898 capability or maximum-range questions; 7,932 causal questions; 7,932 historical-time questions
Reported, quoted, or metalinguistic language	27,762	6.9%	15,864 reported commands and 11,898 quoted or phrase-repetition uses
Past statements	23,796	5.9%	15,864 past movement descriptions; 3,966 past registration statements; 3,966 past stop statements
Future or planned statements	23,796	5.9%	15,864 future movement descriptions; 3,966 future registration statements; 3,966 future stop statements
Conditional or hypothetical expressions	19,830	4.9%	15,864 hypothetical statements and 3,966 condition-triggered movement requests
Other declarative statements	19,830	4.9%	15,864 neutral context statements and 3,966 non-command statements containing a duration
Fragments	47,592	11.8%	Three groups of 15,864: scalar or repetition fragments; standalone continuation or repetition cues; isolated component, direction, or duration slots
Ambiguous or underspecified references	31,728	7.8%	15,864 pronoun or deictic references; 15,864 named components without a complete direction or action
Unresolved self-corrections	23,796	5.9%	Six final-state failure types with 3,966 samples each: pronoun-only intent, unsupported intent, negated intent, multiple intents, component-only fragment, or duration-only fragment
Off-domain command or alternate word sense	39,660	9.8%	19,830 unsupported equipment or room actions; 15,864 command keywords used with another meaning; 3,966 attempts to register a different person's voice

This composition explicitly covers commands, non-commands, negation, questions, statements, quotations, conditional language, future expressions, fragments, ambiguity, multiple commands, conflicts, self-correction, invalid duration expressions, and delayed or periodic requests.

Self-correction coverage

Self-correction language appears in 91,218 samples. In 67,422 samples (73.9%), the final correction restates one complete affirmative target: 51,558 movement samples, 7,932 stop samples, and 7,932 registration samples. This group includes both single corrections and repeated corrections that abandon more than one earlier intent.

The remaining 23,796 samples (26.1%) end in a pronoun-only reference, unsupported direction, negated request, multiple requests, component-only phrase, or duration-only phrase and therefore use <|cmd_none|>. A duration correction produces a movement target only when the corrected duration is attached to a fully restated movement command.

Ambiguity, conflict, and duration boundaries

An utterance with competing actions uses <|cmd_none|> unless an explicit correction marker abandons the earlier action and the final action is complete. Coordinated movements, supported-plus-unsupported combinations, action alternatives, and stop-then-reverse sequences therefore remain within <|cmd_none|>. A single global stop remains a <|cmd_stop|> sample when it names more than one movement to stop.

Positive movement durations are represented as millisecond operands. Zero duration, a bare number without a time unit, a distance in place of time, or two alternative durations does not produce an operand. Delayed execution and periodic scheduling are treated as scheduling requests and use <|cmd_none|>.

Languages and textual style

English and German each contribute 507,648 samples. Both languages cover direct commands, non-commands, short colloquial forms, declarative and interrogative speech, quotations, corrections, and semantic boundary cases. Numerals in spoken text are written as words; numeric operands appear only in the target string.

Data fields

Column	Type and content
audio	Embedded raw speech audio
text	Transcript used to produce the audio
language	English or Deutsch
speaker	Voice identifier
target	Command token with an optional millisecond operand