atasoglu/turkish-emotional-speech-11h

Turkish Emotional Speech (11h)

A fully synthetic Turkish speech dataset with rich labels for expressive TTS and speech emotion recognition (SER). It contains 9,783 clips (11.06 hours, 48 kHz mono FLAC), read by 5 synthetic speakers, in 8 emotions and 4 speaking styles (normal, whisper, speech with laughter, sigh).

Contents

Clips	9,783 (one unique sentence per clip)
Duration	11.06 h; mean 4.1 s (1.8–14.1 s)
Audio	48 kHz, mono, 16-bit FLAC (embedded in parquet)
Speakers	5: 3 female, 2 male
Emotions	neutral, calm, happy, sad, angry, fearful, disgusted, surprised
Styles	normal 8,105 · whisper 805 · speech_laugh 581 · sigh 292
Words	81,364 in total
Distinct words	19,581 word forms (case-insensitive)
Words used only once	11,873 (61 % of the distinct words)

Speakers

speaker_id	Gender	Age group	Clips	Voice
ahenk_f022	female	young adult	1,707	built-in voice of the Ahenk TTS model
v01	female	young adult	1,951	designed: bright and clear
v02	male	adult	2,068	designed: deep, warm baritone
v03	female	middle-aged	1,978	designed: mature, slightly husky
v04	male	young adult	2,079	designed: light, energetic tenor

v01–v04 were created with VoxCPM2 voice design from a text description only, then used as references for cloning. Every speaker reads every emotion (about 195–216 normal-style clips per speaker × emotion cell).

Emotions (all styles): sad 1,441 · calm 1,337 · happy 1,316 · surprised 1,302 · fearful 1,223 · neutral 1,094 · disgusted 1,036 · angry 1,034.

How it was made

1. Text.
   • Five open-weight LLMs wrote short, emotion-bearing Turkish sentences through OpenRouter:
     • Gemma 4 31B (Apache-2.0)
     • DeepSeek V4.1 Flash (MIT)
     • GLM-5.3 Flash (MIT)
     • Hy3 (Apache-2.0)
     • Nex-N2.5 Pro (Apache-2.0)
   • About half of the sentences were written from a random FineWeb-2 Turkish web paragraph given as topical context (text_mode = context; the FineWeb-2 document id is in text_context_fineweb_id). The rest were written without context (baseline).
   • Each sentence also got a random speaker role, form and temperature.
   • Filters removed copies of the context, profanity, sensitive topics, foreign text and ASCII-folded Turkish (ı→i, ğ→g, ş→s).
2. Blind text label. Qwen3.8-27B labelled each sentence's emotion without seeing the target. Only sentences where this label matched the target emotion were kept (text_label_llm).
3. Speech.
   • Normal speech and laughter: Ahenk TTS, a Turkish LoRA on VoxCPM2.
     • Emotion is given as a style instruction (instruction).
     • It speaks in its own voice or clones v01–v04.
     • cfg 2.0 is the medium-expressive setting; 3.0 is the strong one.
   • Whisper and sigh: VoxCPM2 cloning. Each whisper is the least-voiced of 3 seeds.
4. Automatic QC.
   • Whisper large-v3 transcription (asr_text, wer, cer).
   • A voicing check for whispers (voiced_share).
   • A Qwen3-Omni-30B-A3B judge:
     • judge_emotion and judge_p_target: emotion read from the original audio.
     • judge_strength: 1–5 expressiveness, read from low-pass filtered audio so the words are hidden.
5. Filtering. These clips were removed:
   • WER > 0.3 (0.5 for whispers).
   • Whispers that were still voiced.
   • Clips whose last words were lost.
   • Clips with judge_p_target < 0.3 and judge_strength < 3. This rule was calibrated on a 150-clip blind human check.

Quality and known limitations

• Label noise. The label is the intended emotion.
  • A blind human check of 50 random clips, before the final filter, gave 56 % good, 24 % acceptable, 6 % bad and 14 % emotion-mismatch.
  • The automatic judge agrees with the target on 88 % of normal-style clips.
  • For cleaner SER training, filter on judge_emotion == emotion or a judge_p_target threshold.
• Synthetic speech. All voices come from one acoustic model family (VoxCPM2). Models trained only on this data may not generalize to real recordings. Mix it with real speech.
• Pronunciation. Some long vowels in loanwords (e.g. sipariş) are not lengthened. Rare spontaneous sighs or artefacts remain.
• No predefined split. For SER evaluation, hold out a speaker (e.g. v03) or a text LLM.

Fields

Field	Description
audio	48 kHz FLAC
id	clip id (b2_<block>_<speaker>_<text id>_<style>)
text	sentence as written (laugh syllables included for speech_laugh)
text_spoken	text actually sent to the TTS
speaker_id, gender, age_group	speaker
emotion	target emotion (8 classes)
text_intensity	low / medium / high emotional intensity requested from the text LLM (null for neutral)
style	normal, whisper, speech_laugh, sigh
nonverbal	[], ["laugh"] or ["sigh"]
tts_model	voxcpm2+ahenk-lora or voxcpm2
instruction, instruction_lang	style instruction given to the TTS (Turkish when the sentence contains "c", to avoid a c→s mispronunciation)
cfg, steps, seed	generation settings
duration	seconds
text_llm, text_llm_license, text_temperature	sentence generator
text_mode, text_topic, text_speaker_role, text_form	sentence generation prompt facets
text_context_fineweb_id	FineWeb-2 document id of the context paragraph (null for baseline)
text_label_llm	blind text emotion label (Qwen3.8-27B)
asr_text, wer, cer	Whisper large-v3 transcript and error rates
voiced_share	voiced fraction (whispers only)
judge_emotion, judge_p_target, judge_strength	Qwen3-Omni judge outputs (strength: not for neutral / whisper / sigh)

Usage

from datasets import load_dataset

ds = load_dataset("atasoglu/turkish-emotional-speech-11h", split="train")
clean = ds.filter(lambda r: r["judge_emotion"] == r["emotion"])  # stricter SER subset
texts = ds.remove_columns("audio").select_columns(["text", "emotion"])  # text-only emotion classification

Besides speech tasks, the text + emotion pairs can be used for Turkish text emotion classification; every sentence's emotion was confirmed by an independent blind LLM label (text_label_llm).

License

The dataset is released under Apache-2.0 (see LICENSE). All audio and text are synthetic. It was produced with the following third-party models and data, which are acknowledged here.

Speech synthesis

• VoxCPM2 (Apache-2.0) made the speech for all clips. It also designed the voices of speakers v01–v04.
• Ahenk TTS (CC-BY-4.0, by Özgür Güler) is the Turkish LoRA used for normal and laughing speech. Attribution is given as required by CC-BY-4.0. Clips made with it have tts_model = "voxcpm2+ahenk-lora".

Sentence generation (open-weight models, used through OpenRouter)

• Gemma 4 31B (Apache-2.0)
• DeepSeek V4.1 Flash (MIT)
• GLM-5.3 Flash (MIT)
• Hy3 (Apache-2.0)
• Nex-N2.5 Pro (Apache-2.0)

Topical context

• FineWeb-2, Turkish subset tur_Latn (ODC-By 1.0; subject to the Common Crawl terms of use).
  • Paragraphs were used only as prompts. They are not redistributed.
  • Only their FineWeb-2 document ids are listed, in text_context_fineweb_id. Source URLs are not included.

Labelling and quality control

• Qwen3.8-27B wrote the blind text emotion label (text_label_llm).
• Qwen3-Omni-30B-A3B-Instruct (Apache-2.0; AWQ 4-bit build cyankiwi/Qwen3-Omni-30B-A3B-Instruct-AWQ-4bit) produced the judge_* fields.
• Whisper large-v3 (MIT; via Systran/faster-whisper-large-v3) produced asr_text, wer and cer.
• The hermitdave/FrequencyWords Turkish list drove the ASCII-folding detector. It was used only to filter text.