TartarusXXX/mixed-language-detection-english-accented-vc

Mixed-Language Speech Detection Pilot

This dataset is a 6,000-clip binary audio-classification pilot for detecting whether an utterance contains one language (label = 0) or more than one language (label = 1). It covers Turkish (tur), Northern Kurdish/Kurmanji (kmr), Central Kurdish/Sorani (ckb), Arabic (ara), Persian (fas), and English (eng).

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. Cloned voice references, source recordings, and text pools 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. Twenty distinct synthetic reference voices were first created with OmniVoice voice design, then used through the voice-cloning interface. Twelve references are exclusive to train, four to validation, and four to test. No real person's voice was cloned.

For segment-level mixed clips, 80% use one cloned voice across all languages and 20% alternate between two cloned voices. Switch patterns are approximately 70% A→B, 25% A→B→A, and 5% A→B→A→B. Segment boundaries in segments reflect the assembled audio before length-preserving full-clip augmentation.

Single-call mixed generation uses language-agnostic conditioning and should be treated as the experimental subset described by construction = "single_call_tts". It is intended primarily for clip-level evaluation.

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 in each real sample. No 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: unique ISO 639-3 language codes present
• language_sequence: ordered span languages
• speaker_condition: same or different voice condition
• segments: start/end time, language, text, and provenance per span; null for single-call generation
• text: complete input text for single-call generation; null otherwise
• audio_condition and augmentation_detail: full-clip audio treatment
• reference_voice_ids, source_ids, source_corpora: provenance
• qc_*: deterministic file-integrity and signal checks

Validation and limitations

Every published file is checked for successful decoding, nonzero duration, finite samples, sample rate, clipping, annotation consistency, exact class and construction counts, split isolation, source reuse, and distribution targets. ASR and audio-language-identification scores are intentionally not included. Consequently, intended languages—especially in the experimental single-call subset—are annotations of the generation request rather than guarantees from an independent acoustic recognizer. Human listening is recommended before using this pilot as a gold evaluation set.

Licensing and attribution

The derived dataset is released under CC BY 4.0. Users should also cite the upstream FLEURS, KMR-Tatoeba, LibriSpeech, and OmniVoice resources when using the corresponding subsets. Text and source identifiers are included to support auditing and attribution.