TartarusXXX/mixed-language-detection-pilot-complete-sentences

Mixed-Language Speech Detection Pilot — Complete Sentences

This is the complete-sentence revision of a 6,000-clip binary audio-classification pilot. label = 0 denotes one intended language and label = 1 denotes more than one intended language. The covered languages are Turkish (tur), Northern Kurdish/Kurmanji (kmr), Central Kurdish/Sorani (ckb), Arabic (ara), Persian (fas), and English (eng).

What changed

Earlier generation forced source transcripts into arbitrary duration buckets by retaining only their first N words. That could create punctuated fragments such as Göreceli ulaşılmazlığıyla bir. This revision prohibits token slicing. Every synthetic span is an exact, whitespace-normalized copy of a complete source transcript unit.

The deterministic acceptance rules require terminal punctuation, 4–28 words, enough lexical content, balanced brackets, no URL/placeholder/ellipsis markers, and no language-specific continuation word at the end. Target TTS duration is derived from the accepted sentence length (1.8–13.5 seconds per segment), not the other way around. Single-call inputs are exact joins of accepted units and expose their components in prompt_units.

There are 1,752 accepted source units. The build audits all 12,975 planned text units—including real-speech transcripts—against their original source IDs. Exactly 25% of each synthetic monolingual construction is deliberately selected from sentences containing a named entity, number, borrowed word, or technical term. This prevents a detector from treating technical vocabulary as automatic evidence of a language switch.

Native voice references

Synthetic speech is cloned from 36 real, native-language reference recordings: two references for every language in every dataset split. Turkish, Sorani, Arabic, Persian, and English references come from Google FLEURS. Kurmanji references come from FLEURS-Kobani (aranemini/northern-kurdish-fleurs).

For monolingual synthesis, the reference language always matches the generated language. For mixed clips using different voices, each language uses a matching native reference. For mixed clips using one voice, the reference is drawn from one of the two languages; this unavoidable cross-language cloning is disclosed by reference_languages and reference_strategy.

Reference recordings are never included as real-concatenation segments and are isolated by dataset split. Official FLEURS does not publish person-level speaker IDs, so person-level separation beyond recording isolation cannot be guaranteed for those references. FLEURS-Kobani speaker IDs are separated across splits.

Dataset composition

Construction	Mixed	Monolingual	Total
Single-call OmniVoice	500	500	1,000
Segment-level OmniVoice	1,500	1,500	3,000
Real-speech concatenation	1,000	1,000	2,000
Total	3,000	3,000	6,000

The split is 80/10/10: 4,800 train, 600 validation, and 600 test clips. Text, reference recordings, and real source recordings are isolated by split.

Audio and generation

• 16 kHz mono FLAC
• 60% clean
• 20% synthetic colored background noise
• 10% telephone/radio filtering
• 10% compression or short-room reverberation

Synthetic clips were generated with k2-fsa/OmniVoice using FLEURS-family voice-cloning references. For segment-level mixed clips, 80% use one cloned voice across all languages and 20% use language-matched voices. Switch patterns are approximately 70% A→B, 25% A→B→A, and 5% A→B→A→B. segments contains the measured post-assembly boundaries.

Single-call mixed generation uses one reference for the whole utterance and has no reliable timestamps. Its segments value is null, text is the exact model input, and prompt_units records each complete component and its language.

Real speech

Real concatenations use recordings from:

• Google FLEURS (google/fleurs), CC BY 4.0
• FLEURS-Kobani (aranemini/northern-kurdish-fleurs), CC BY 4.0
• KMR-Tatoeba (aranemini/kmr-tatoeba), CC BY 4.0

No real source recording is used more than three times or shared across dataset splits. Half of the real joins are same-corpus and half different-corpus for each label.

Important fields

• audio, id, label, construction
• languages, language_sequence, speaker_condition
• segments: measured boundaries, complete text, text provenance, and voice provenance for segment TTS and real concatenations
• text and prompt_units: exact single-call prompt and its complete units
• text_id, text_source_id, text_source_collection, text_category, and complete_source_unit: sentence-quality evidence
• reference_* and reference_strategy: cloning provenance
• audio_condition, augmentation_detail, source_ids, and source_corpora
• qc_*: deterministic file-integrity and signal checks

Validation and limitations

The release validator checks every published audio file, distribution, split, source-use limit, reference policy, annotation boundary, and text unit. A synthetic unit fails publication unless it matches the accepted source catalog exactly; a real segment fails unless it matches its source recording transcript.

The source material consists of human-authored or human-translated corpus transcripts. Some translations may still sound formal or awkward, but generated token fragments are no longer permitted.

ASR and audio-language-identification models are intentionally not used. Language annotations describe construction inputs rather than independent acoustic verification. Native references reduce English-accent bias but cannot guarantee a native accent when one voice is asked to speak another language. Human listening is recommended before using this pilot as a gold evaluation set. The real speakers used as cloning references are intended for language-detection research, not imitation or impersonation.

Licensing and attribution

The derived dataset is released under CC BY 4.0. Users should cite Google FLEURS, FLEURS-Kobani, KMR-Tatoeba, and OmniVoice for their corresponding portions. Source identifiers are retained for audit and attribution.