grushaaaaa/kannada-emotional-tts

Grusha Kannada Emotional TTS

A single-speaker Kannada (ಕನ್ನಡ) speech dataset for text-to-speech (TTS) and expressive / emotional speech synthesis, recorded by a single female speaker (grusha_kannada). Every utterance is labelled with one of four emotions — neutral, happy, sad, angry — making the corpus suitable for training expressive and emotion-controllable TTS models, as well as speech-emotion classification.

Dataset at a glance

Language	Kannada (kn)
Speaker	Single female speaker (grusha_kannada)
Utterances	1,135
Total duration	~1.92 hours
Sample rate	16 kHz, mono, 16-bit PCM
Emotions	neutral, happy, sad, angry
Domain	Conversational — healthcare, banking/customer-care, everyday dialogue
Script	Kannada, with occasional inline English/code-mixed terms

Per-emotion breakdown

Emotion	Utterances	Duration
neutral	564	0.98 h
happy	209	0.42 h
sad	278	0.35 h
angry	84	0.17 h
Total	1,135	~1.92 h

Usage

from datasets import load_dataset

ds = load_dataset("grushaaaaa/kannada-emotional-tts", split="train")
print(ds)

ex = ds[0]
print(ex["text"], "|", ex["emotion"])
print(ex["audio"]["sampling_rate"], len(ex["audio"]["array"]))

# Filter to a single emotion
happy = ds.filter(lambda r: r["emotion"] == "happy")

Data fields

Field	Type	Description
audio	Audio(16 kHz)	Mono waveform, 16-bit PCM WAV
text	string	Kannada transcript of the utterance
emotion	ClassLabel	One of neutral, happy, sad, angry
duration	float32	Clip duration in seconds
speaker_id	string	Always grusha_kannada (single speaker)
language	string	Language code (kn)
source_recording	string	Source session the clip was segmented from
snr_db	float32	Estimated signal-to-noise ratio (dB)
dnsmos_ovrl	float32	DNSMOS overall MOS quality estimate
silence_ratio	float32	Fraction of the clip that is silence
clipping	float32	Clipping metric (0 = none)
rms	float32	RMS loudness of the clip

Collection & processing

• Recorded by a single native Kannada speaker across multiple sessions.
• Raw sessions were speech-enhanced, then segmented into single-utterance clips.
• Each clip was transcribed and passed through an automatic QA pass that computes SNR, DNSMOS, silence ratio, clipping and RMS (retained in the metadata so downstream users can filter on quality).
• Clips were grouped by emotional delivery into the four emotion classes.
• All audio is resampled to a uniform 16 kHz mono and stored as 16-bit PCM.

Recommended use

• Training / fine-tuning single-speaker Kannada TTS models.
• Emotion-controllable or expressive TTS (condition on the emotion label).
• Speech-emotion recognition for Kannada.
• Quality filtering: e.g. keep snr_db > 10 and dnsmos_ovrl > 2.5 for the cleanest subset.

Limitations & considerations

• Single speaker — models trained on this alone will not generalise across voices.
• Class imbalance — neutral dominates; angry is the smallest class. Consider re-weighting or resampling for balanced training.
• Code-mixing — some transcripts contain inline English/romanised terms reflecting natural conversational Kannada; text normalisation may be needed for some pipelines.
• Emotion labels reflect the speaker's intended delivery per session, not per-frame affect annotation.

License

Released under CC BY-NC 4.0 (non-commercial, attribution). Please contact the dataset owner for commercial licensing.

Citation

@misc{grusha_kannada_emotional_tts_2026,
  title  = {Grusha Kannada Emotional TTS},
  author = {Grusha},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/datasets/grushaaaaa/kannada-emotional-tts}}
}