SherTheCoder/TeamFolklore_UthoKikkar

Utho Kikkar

Synthetic and Real-world speech for teaching a tiny chip one wake word, in Punjabi and Hindi.
Voices saying "Utho Kikkar", and many more saying things that only sound like it.

What this is

Utho Kikkar (ਉੱਠੋ ਕਿੱਕਰ · उठो किक्कर) means "wake up, Kikkar". Kikkar is the babul tree you see along canal banks in Punjab. It is also the name of our project: a microcontroller that listens for this phrase on its own, opens a flower on its screen when it hears it, and only then streams your command to the cloud.

This dataset is what we trained that listener on. A wake-word model has two jobs. It has to fire when someone says the phrase, and it has to stay quiet for everything else, most of all for phrases that sound nearly the same. So every clip carries one of two labels:

Label	Meaning	Example
1	The wake word	ਉੱਠੋ ਕਿੱਕਰ · उठो किक्कर
0	Anything else, including near-misses	ਉੱਠੋ ਕਿਰਨ ("Utho Kiran")

The near-misses are the important part. Utho Kiran is only a couple of sounds away from Utho Kikkar.

How synthetic

AI4Bharat's Indic Parler-TTS speaks Indian languages in many named voices, and each one is steered with a plain English description like "Rohit speaks in an expressive, slightly higher-pitched voice at a normal pace". That let us cover a wide range of voices, pitches and speaking speeds to augment our real dataset.

How synthetic data was made

1. Phrase. Either the wake word, or something the model must not react to.
2. Voice prompt. A named speaker and a description of how they sound.
3. Generation. Indic Parler-TTS at a fixed model revision. The seed is saved with every clip, so any clip can be traced back to exactly how it was made.
4. Clean up. Silence is trimmed from both ends and the audio is resampled to 16 kHz, the rate the Kikkar board listens at. Volume is normalized, and a maximum duration of three seconds is set.
5. Checks. Manual checks on every generated clip, along with automatic screening using IndicConformer CTC/RNNT, Meta Omnilingual ASR v2 and speech-boundary checks.
6. Record. Every clip gets a JSON record with its text, voice prompt, seeds, lengths and a SHA-256 of the audio.

One record

Each split folder has a metadata.jsonl with one line per clip. This is a real one:

{
  "sample_id": "clip_06fe807f04b0e8c4a2041b30",
  "text": "ਉੱਠੋ ਕਿੱਕਰ।",
  "label": 1,
  "split": "train",
  "speaker": "Rohit",
  "seed": 117342335,
  "attempt": 220,
  "policy": "independent-asr-v2",
  "sample_rate": 16000,
  "duration_seconds": 1.35125,
  "gain_db": -3.2536043204883636,
  "peak_dbfs": -9.632484820128152,
  "trim_start_seconds": 0.1,
  "trim_end_seconds": 1.45125,
  "pcm_sha256": "43706263cf86cc860020f73d97789b0fd030d88ab9166f71ae5bcc233f01d43f",
  "usable": true,
  "sha256": "573e755a28c56044b2fa4faab25441d4c86f2eeb23fe8ade68fe4a1df228d84a",
  "accepted": true,
  "reason": "independent_transcript_consensus",
  "transcripts": {
    "indic_ctc": "ਉੱਠੋ ਕਿੱਕਰ",
    "indic_rnnt": "ਉੱਠੋ ਕਿੱਕਰ",
    "omni": "ਉੱਠੋ ਕਿੱਕਰ"
  },
  "temporal_ok": true
}

Fields

Field	Type	Description
sample_id	string	Unique recording identifier
text	string	Intended text supplied to TTS
label	int	1 = wake word; 0 = negative
split	string	Assigned split: train, validation or test
speaker	string	Named Indic Parler-TTS voice
seed	int	Random seed used for generation
attempt	int	Generation-attempt index
policy	string	Automatic screening policy version
sample_rate	int	Sampling rate in Hz; 16000 for this dataset
duration_seconds	float	Cleaned clip duration in seconds
gain_db	float	Normalization gain; positive amplifies, negative attenuates
peak_dbfs	float	Peak audio level after normalization
trim_start_seconds	float	Start of the retained interval in the source audio
trim_end_seconds	float	End of the retained interval in the source audio
pcm_sha256	string	PCM audio checksum used for duplicate detection
usable	bool	Whether basic audio cleanup checks passed
sha256	string	SHA-256 checksum of the saved WAV file
accepted	bool	Whether automatic dataset acceptance checks passed
reason	string	Reason recorded for the acceptance decision
transcripts	object	Recognized text from IndicConformer’s CTC and RNNT decoders (indic_ctc, indic_rnnt) and Meta Omnilingual ASR (omni)
temporal_ok	bool	Whether speech-boundary and unexplained-tail checks passed

Using it

from datasets import load_dataset

ds = load_dataset("SherTheCoder/TeamFolklore_UthoKikkar")

clip = ds["train"][0]
clip["audio"]["array"]          # the waveform, 16 kHz
clip["label"], clip["text"]

Only the wake word:

wake = ds["train"].filter(lambda row: row["label"] == 1)

Just the files, without datasets:

huggingface-cli download SherTheCoder/TeamFolklore_UthoKikkar \
  --repo-type dataset --local-dir utho-kikkar

If you are training for the Kikkar board, compute features the way the firmware does: 16 kHz audio, a 30 ms Hann window every 10 ms, a 512-point FFT, 40 mel bands from 20 Hz to 7.6 kHz, the natural log floored at -80 dB, and a running mean per band subtracted with a 3 s time constant. The model trained on this data is SherTheCoder/TeamFolklore_Model, and its card has the reference code.

Licence

The dataset is released under CC BY 4.0. The synthetic audio was generated with Indic Parler-TTS by AI4Bharat, which is released under the Apache 2.0 licence.

Citation

@misc{teamfolklore2026uthokikkar,
  title        = {Utho Kikkar: synthetic wake-word speech in Punjabi and Hindi},
  author       = {{Team Folklore}},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/SherTheCoder/TeamFolklore_UthoKikkar}}
}

The voices come from Parler-TTS, so please cite it too:

@misc{lyth2024natural,
  title         = {Natural language guidance of high-fidelity text-to-speech with synthetic annotations},
  author        = {Dan Lyth and Simon King},
  year          = {2024},
  eprint        = {2402.01912},
  archivePrefix = {arXiv},
  primaryClass  = {cs.SD}
}

Made by

Team Folklore.