This dataset is curated and maintained by Zhagaram Labs and requires explicit manual review and whitelisting prior to access. Please complete the access request form below with your organization details and intended use case. Approved users and organizations will be granted access by the Zhagaram Labs team.

Zhagaram Labs Tamil & Tanglish Conversational Speech Corpus

Dataset Overview

The Zhagaram Labs Tamil & Tanglish Conversational Speech Corpus is a large-scale, studio-grade spontaneous speech dataset designed for training and evaluating Automatic Speech Recognition (ASR), Text-to-Speech (TTS), Speech-to-Speech, and Spoken Language Models across both Pure/Standard Spoken Tamil (tamil) and Code-Switched Tamil–English (tanglish).

Unlike traditional read-speech corpora recorded from scripted prompts, this corpus captures natural, long-form multi-speaker studio conversations, interviews, panel discussions, and domain-diverse monologues. It reflects authentic spoken prosody, natural discourse markers, conversational turn-taking, and real-world bilingual code-switching patterns spoken across Tamil Nadu and the global Tamil diaspora.

Access Policy: This dataset is gated and explicitly whitelisted. All access requests are manually reviewed and approved by Zhagaram Labs.

Corpus Summary Statistics

Metric	Pure Tamil (`tamil`)	Tanglish (`tanglish`)	Combined Corpus
Total Clean Speech Duration	110.00 hrs	100.00 hrs	210.00 hrs
Total Utterance Segments	35,000	32,000	67,000
Unique Long-Form Conversations	130	120	250
Target Utterance Window	4.0s – 15.0s	4.0s – 15.0s	4.0s – 15.0s (avg 11.2s)
Audio Format	16 kHz, 16-bit Mono PCM WAV	16 kHz, 16-bit Mono PCM WAV	16 kHz, 16-bit Mono PCM WAV

Two-Stage Audio Processing & Curation Pipeline

Every long-form studio recording undergoes a rigorous, multi-stage acoustic cleaning and segmentation pipeline prior to inclusion:

flowchart LR
    A["Raw Long-Form Studio Audio"] --> B["Stage 1: Vocal Stem Isolation (Demucs htdemucs)"]
    B --> C["Stage 2: 16kHz Mono Standardization"]
    C --> D["Stage 3: Silero VAD Speech Boundary Detection"]
    D --> E["Stage 4: Smart Conversational Utterance Packing (4s - 15s)"]
    E --> F["Stage 5: RMS Energy Gating & Silence Rejection"]
    F --> G["Stage 6: Hybrid Acoustic-Linguistic Tagging (Tamil vs. Tanglish)"]

1. Vocal Isolation & Background Suppression

Hybrid Transformer Demucs (htdemucs): Studio intros, outros, and underlying background music beds are stripped using two-stem vocal separation (--two-stems=vocals) where applicable, isolating clean human voice tracks.

2. Neural Voice Activity Detection & Utterance Packing

Silero VAD Segmentation: Speech boundaries are detected at 16,000 Hz using a confidence threshold of 0.5, a minimum silence split duration of 400 ms, and 150 ms boundary padding to prevent word-edge clipping.
Smart Conversational Utterance Packing: Raw VAD detectors frequently fragment natural speech on brief inhalations (producing 1-second fragments). Our utterance packer merges adjacent speech intervals across natural conversational pauses (<= 1.2 seconds) into cohesive 4.0 to 15.0 second utterances while splitting extended monologues at the cleanest acoustic silence boundary.
RMS Energy Verification: Segments with root-mean-square (RMS) amplitude below 0.005 are discarded to eliminate near-silent room tone or false-positive breath triggers.

3. Hybrid Acoustic-Linguistic Tagging (`tamil` vs. `tanglish`)

Each conversation and its constituent utterances are tagged into one of two linguistic register subsets using a hybrid acoustic and lexical probing pipeline:

Multi-Segment Acoustic Language Identification: Evenly-spaced interior segments across each conversation are probed with neural acoustic language identification to compute mean Tamil acoustic probability (acoustic_ta_prob), English acoustic probability (acoustic_en_prob), relative English share, and 75th-percentile code-switch intensity.
Lexical Code-Switch Ratio: Representative mid-conversation segments are decoded unconstrained to measure the ratio of embedded English lexical items within Tamil syntactic frames.
Composite tanglish_score (0.0 – 1.0):
- tamil (tanglish_score < 0.42): Predominantly pure spoken Tamil with minimal or only incidental English loanwords.
- tanglish (tanglish_score >= 0.42): Conversational Tamil–English code-switching (intra-sentential and inter-sentential switching typical of urban, technical, business, and contemporary conversational registers).
- Non-Tamil outlier recordings are automatically filtered out.

Dataset Structure & Schema

The dataset provides two splits (tamil and tanglish) under the default configuration, as well as dedicated tamil and tanglish configurations.

Data Fields

Column	Type	Description
`segment_id`	`string`	Unique anonymized identifier for the audio segment (e.g., `conv_0001_seg_0001`).
`conversation_id`	`string`	Anonymized identifier grouping segments from the same recording session (`conv_0001`).
`audio`	`Audio(sampling_rate=16000)`	16 kHz, 16-bit mono PCM WAV audio waveform array and metadata.
`duration_sec`	`float32`	Duration of the audio segment in seconds (`4.0` to `15.0`).
`start_sec`	`float32`	Start timestamp (in seconds) within the original recording session.
`end_sec`	`float32`	End timestamp (in seconds) within the original recording session.
`language_style`	`string`	Linguistic register tag: `"tamil"` (Pure/Standard Spoken Tamil) or `"tanglish"` (Code-Switched Tamil–English).
`tanglish_score`	`float32`	Continuous code-switching index in `[0.0, 1.0]`, where higher values indicate heavier Tamil–English mixing.
`acoustic_ta_prob`	`float32`	Mean neural acoustic language identification probability for Tamil (`ta`).
`acoustic_en_prob`	`float32`	Mean neural acoustic language identification probability for English (`en`).

Usage

Once your Hugging Face account has been whitelisted and approved for access:

from datasets import load_dataset

# Load both splits ('tamil' and 'tanglish')
ds = load_dataset("zhagaramlabs/tamil-conversational-speech", token=True)

print(ds["tamil"][0])
print(ds["tanglish"][0])

# Or stream directly without downloading the full corpus
stream_ds = load_dataset(
    "zhagaramlabs/tamil-conversational-speech",
    split="tamil",
    streaming=True,
    token=True
)
for sample in stream_ds.take(5):
    audio_array = sample["audio"]["array"]
    sr = sample["audio"]["sampling_rate"]
    print(sample["segment_id"], sample["language_style"], sample["duration_sec"], sr)

Intended Uses & Applications

Bilingual & Code-Switched ASR: Pretraining and fine-tuning Whisper, Conformer, and transducer architectures to handle seamless Tamil–English code-switching without hallucinating or transliterating English technical terms.
Conversational TTS & Voice Cloning: Training expressive, natural-prosody Tamil and Tanglish speech synthesis models on clean 4–15s utterances.
Spoken Language & Audio Foundation Models: Pretraining discrete audio tokenizers and speech-native multimodal LLMs on spontaneous Dravidian conversational speech.
Linguistic Code-Switching Research: Studying phonological adaptation, discourse markers, and turn-taking in spontaneous Tamil–English bilingual speech.

Licensing & Access Governance

Access to this dataset is restricted to explicitly whitelisted researchers, developers, and partner organizations approved by Zhagaram Labs. Approved users may not redistribute, mirror, or publicly re-host the raw audio files.

Downloads last month: -

Total file size:

12 kB

You need to agree to share your Hugging Face information to access this dataset

Explicit Whitelist & Access Approval Required