burak-ozenc/dawn-chorus-codec-labels

Dawn Chorus EN - Codec Labels

Sidecar dataset for ai-coustics/dawn_chorus_en.

Adds a codec_guess column, applying hypotheses to classify the audio source type(GSM, WhatsApp, Telegram) which is not present in current dataset, depending on spectral analysis of speech channel on original dataset.

Since audio source type distribution given in the actual dataset(67% GSM, 16.5% WhatsApp, 16.5% Telegram), this classification is unsupervised, guided by the known prior distribution.

Motivation

The dawn_chorus_en dataset contains 3 different transmission channels without specifying the source type(GSM, WhatsApp, Telegram). Enhancement models (DeepFilterNet, NoiseReduce) report aggregated metrics (SI-SDR, PESQ, WER) that obscure codec-specific performance differences. So, I decided to run some hypotheses on the actual dataset.

Method

All audio in dawn_chorus_en is resampled to 16kHz, which destroys native sample rate. This classification uses two spectral features to identify what is lost during preprocessing:

1. bw_99 - the frequency below which 99% of the signal's power falls, computed via Welch PSD. GSM has a hard around 3.4kHz codec ceiling that remains detectable even after upsampling.

2. spectral_slope - slope of a log-log linear fit of PSD above 1kHz. Steeper (more negative) = sharper rolloff = heavier codec compression. Used to separate WhatsApp from Telegram within the wide-bandwidth group.

Decision rule

def classify_codec(row):
    if row['bw_99'] < 4472:
        return 'GSM'
    if row['spectral_slope'] < -0.45:
        return 'WhatsApp'
    return 'Telegram'

GSM was a hard-edged case, it has 3.4 kHz frequency ceiling. I run on a 200(to be able to handle error rate better) sample, to match 67% of GSM rate, where to set the cutoff value using scipy.optimize.minimize_scalar.

4000 Hz - 110 GSM - too few, distance = 24
5000 Hz - 160 GSM - too many, distance = 26
4472 Hz - 134 GSM - closest to 134, distance = 0 

Results

codec_guess	count	proportion	expected
GSM	299	66.4%	67.0%
WhatsApp	88	19.6%	16.5%
Telegram	63	14.0%	16.5%

GSM classification is highly reliable (0.6% off expected). WhatsApp and Telegram are slightly misclassified into each other (around 10-15 samples) due to spectral slope overlap in the ambiguous zone after resampling.

Key findings from cross-tab analysis

• WhatsApp is 64% machine-generated speech - the highest synthetic voice ratio across all three codecs. GSM and Telegram are 70-83% human.
• Telegram has 35% narrative content vs around 7% for GSM/WhatsApp - reflecting how each platform was used during collection (monologue sharing vs phone calls).
• Codec is speaker-specific - certain speakers appear almost exclusively in one codec. Speaker identity and codec are correlated, meaning aggregated benchmark metrics cannot isolate codec effect from speaker characteristics.

Schema

column	type	description
id	string	original sample id from dawn_chorus_en
speaker_id	string	speaker identifier
language	string	always en
conversation_type	string	interactive or narrative
speech_source	string	human or machine
index	int64	original sample index
bw_95	float	frequency below which 95% of power falls (Hz)
bw_99	float	frequency below which 99% of power falls (Hz)
spectral_slope	float	log-log PSD slope above 1kHz
ceiling_ratio	float	energy ratio 4k-8k band / 3k-4k band
vad_rate	float	silence transitions per second
spectral_flatness	float	mean spectral flatness
mfcc_hi_std	float	std of MFCC coefficients 9-13
r_0_3k	float	normalized energy share 0-3kHz
r_3k_4k	float	normalized energy share 3-4kHz
r_4k_8k	float	normalized energy share 4-8kHz
r_8k_p	float	normalized energy share 8kHz+
codec_guess	string	GSM, WhatsApp, or Telegram

Usage

from datasets import load_dataset

labels = load_dataset("burak-ozenc/dawn-chorus-codec-labels", split="train")
df = labels.to_pandas()

# join with original dataset by id
# df.merge(your_results_df, on='id')

# filter by codec
gsm_only = df[df['codec_guess'] == 'GSM']

Limitations

• Classification is semi-supervised - no ground truth labels exist in the original dataset
• WhatsApp/Telegram boundary is soft due to spectral slope overlap after 16kHz resampling
• All features extracted from the speech channel (clean), not the mix channel
• Eval split only (450 samples)

Citation

@dataset{dawn_chorus_en,
  title        = {dawn_chorus_en: An evaluation dataset for accurate foreground speaker transcription},
  author       = {Leonardo Nerini and Butch Warns and Joschka Wohlgemuth and Luis Küffner and Théo Fuhrmann},
  year         = {2026},
  publisher    = {ai-coustics GmbH},
  license      = {CC BY-NC 4.0},
  url          = {https://ai-coustics.com}
}