TartarusXXX/mixed-language-detection-pilot-fleurs-voices

Mixed-Language Speech Detection Pilot — Native FLEURS Voices

This is the native-reference 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 in this revision

Synthetic speech is cloned from 36 real, native-language reference recordings instead of English voice-designed references. There are 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), a native-speaker extension of the FLEURS benchmark.

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 and both directions are balanced within each language pair. The fields reference_languages and reference_strategy make this choice explicit.

The reference recordings are never included as real-concatenation source segments. Reference IDs and recordings are isolated by dataset split. FLEURS does not publish person-level speaker IDs for the five official configurations, so person-level separation beyond recording isolation cannot be independently 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

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

Augmentation is applied after complete clip assembly and has the same distribution for both labels.

Synthetic speech

Synthetic clips were generated with k2-fsa/OmniVoice through its voice-cloning interface. Unlike the earlier synthetic-reference revision, this dataset uses real FLEURS-family recordings as the cloning references. The reference source, language, transcript, and dataset split are retained in the generation manifest used to create the dataset.

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. Segment boundaries in segments describe the assembled audio before length-preserving full-clip augmentation.

Single-call mixed generation necessarily uses one reference for the whole utterance and has no reliable segment timestamps. Its segments value is null, while text contains the complete generation input.

Real speech

Real concatenations use recordings from:

• Google FLEURS (google/fleurs), CC BY 4.0
• KMR-Tatoeba (aranemini/kmr-tatoeba), CC BY 4.0
• LibriSpeech dev-clean, CC BY 4.0

Source recording IDs and corpora are retained. 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, independently for mixed and monolingual labels.

Fields

• audio: viewer-decoded audio from file_name
• id: stable sample identifier
• label: 0 for monolingual, 1 for mixed-language
• construction: single_call_tts, segment_tts, or real_concatenation
• languages and language_sequence: intended language set and ordered spans
• speaker_condition: same- or different-voice construction
• segments: boundaries, language, text, native reference language, and provenance per span; null for single-call generation
• text: complete input text for single-call generation; null otherwise
• reference_voice_ids, reference_languages, reference_source_ids, reference_source_repositories, reference_source_splits, and reference_strategy: cloning provenance and cross-language-reference policy
• audio_condition and augmentation_detail: full-clip audio treatment
• source_ids, source_corpora, corpus_condition: real-speech provenance
• qc_*: deterministic file-integrity and signal checks

Validation and limitations

Every published file is checked for decoding, nonzero duration, finite samples, sample rate, clipping, annotation consistency, exact distributions, split isolation, source reuse, reference-language policy, and source/reference separation.

ASR and audio-language-identification models are intentionally not used. Intended language annotations therefore describe construction inputs rather than independent acoustic verification. A native reference can reduce the English-accent bias but does not guarantee a native accent when that voice is asked to speak another language. Human listening remains recommended before using this pilot as a gold evaluation set.

This revision uses real speakers as voice-cloning references. It is intended for language-detection research, not speaker 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, LibriSpeech, and OmniVoice when using their corresponding portions. Source identifiers are retained for audit and attribution.