audio dict | text stringlengths 15 182 | speaker_name stringclasses 1
value |
|---|---|---|
{"bytes":"UklGRiQYBgBXQVZFZm10IBAAAAABAAEAwF0AAIC7AAACABAAZGF0YQAYBgACAAcAEAAJAP7/+P/z//b/+f/1//H/8f(...TRUNCATED) | "वाई सर, मुझे पता नहीं था कि warranty का time limit 30 (...TRUNCATED) | munni |
{"bytes":"UklGRojOBABXQVZFZm10IBAAAAABAAEAwF0AAIC7AAACABAAZGF0YWTOBAABAAIAAAAAAAAAAgABAAAAAAABAAAA//(...TRUNCATED) | "मुझे repair के लिए फिरसे call करेंगे (सात एक प(...TRUNCATED) | munni |
{"bytes":"UklGRqRcBQBXQVZFZm10IBAAAAABAAEAwF0AAIC7AAACABAAZGF0YYBcBQAHAAUABwAKAAMAAQD9//z/+//6//r/+v(...TRUNCATED) | "हाँ, मैं नहीं आ सकता कल पॉलिसी warranty के लि(...TRUNCATED) | munni |
{"bytes":"UklGRqB4BQBXQVZFZm10IBAAAAABAAEAwF0AAIC7AAACABAAZGF0YXx4BQAAAAAAAAABAAAAAQAAAAEAAQAAAAEAAA(...TRUNCATED) | "मैं call करूंगा छह चार चार चार एक पाँच ती(...TRUNCATED) | munni |
{"bytes":"UklGRqTlAwBXQVZFZm10IBAAAAABAAEAwF0AAIC7AAACABAAZGF0YYDlAwAAAAEAAgABAAAAAAD//////v////7///(...TRUNCATED) | "अबे यार, हमारा plan finalize करो ना loan लेने का, स(...TRUNCATED) | munni |
{"bytes":"UklGRmYFBgBXQVZFZm10IBAAAAABAAEAwF0AAIC7AAACABAAZGF0YUIFBgD///7//v///wAAAAD///7//f///wAAAQ(...TRUNCATED) | "वही 28 लोगों का समूह, प्लान्स discuss करते कर(...TRUNCATED) | munni |
{"bytes":"UklGRrioBQBXQVZFZm10IBAAAAABAAEAwF0AAIC7AAACABAAZGF0YZSoBQACAAMAAgAEAAUABAADAAMABAADAAQABA(...TRUNCATED) | "कहो ना, क्या करें?। Phone: एक दो तीन चार पाँ(...TRUNCATED) | munni |
{"bytes":"UklGRiKfBgBXQVZFZm10IBAAAAABAAEAwF0AAIC7AAACABAAZGF0Yf6eBgAAAAEAAAABAAEAAAAAAP/////+/////v(...TRUNCATED) | "वो एक महिला बोली, अरे, मुझे लगता है कि य(...TRUNCATED) | munni |
{"bytes":"UklGRtLZBwBXQVZFZm10IBAAAAABAAEAwF0AAIC7AAACABAAZGF0Ya7ZBwAAAAEAAAABAAEAAQAAAAAAAQAAAAEAAA(...TRUNCATED) | "उसका दोस्त मुस्कुराया और बोला, हां, बि(...TRUNCATED) | munni |
{"bytes":"UklGRuQrBQBXQVZFZm10IBAAAAABAAEAwF0AAIC7AAACABAAZGF0YcArBQAt+3P7X/2z/Ef9Wv03/3QBdgU5CNsJAw(...TRUNCATED) | "आरे भाई, मेरे फोन की warranty खत्म हो गई है न, (...TRUNCATED) | munni |
Munni — Hindi/English Synthetic TTS Voice Dataset
Single-speaker, code-mixed Hindi (Devanagari) + English speech dataset built for XTTS-v2 fine-tuning. Domain is call-center / customer-support style dialogue (warranty, billing, appointment scheduling), with scripted placeholder phone numbers spoken digit-by-digit.
- 8,796 clips / 18.16 hours, 24 kHz mono 16-bit PCM
- Split: train 8,621 / validation 175 (98/2, seed 42)
- Duration per clip: mean 7.43s, min 3.18s, max 11.00s
- Single speaker:
munni
Source
Derived from a 25.24-hour, 5,792-clip source recording (44.1kHz mono, code-mixed Devanagari + English), processed down to shorter, cleaner segments suitable for XTTS-v2 conditioning windows.
Pipeline
| Stage | Tool | Result |
|---|---|---|
| Forced alignment | ctc-forced-aligner, MMS-300M, --romanize, lang=hin |
5,792 aligned, 0 failures |
| Segmentation | punctuation-priority cuts | 12,150 segments |
| QA filter | whisper-hindi-large-v2 CER | 8,796 kept, 3,354 rejected |
Segmentation: target range 3.0–11.0s (XTTS drops clips >11.6s; ~3s minimum for conditioning). Cuts prioritized at sentence-final punctuation (। ? ! .), then commas, then largest silence gap, placed at the midpoint of the inter-word silence, with ≤150ms padding per side.
QA filter: CER computed on raw strings (no script normalization) against
vasista22/whisper-hindi-large-v2 transcriptions. Threshold set at the p90 of
the CER distribution (0.5279) rather than a fixed cutoff, since absolute CER
is inflated by whisper-hindi's tendency to transliterate English into
Devanagari and to drop punctuation — the filter is relative, not absolute.
Number handling
Transcripts spell out digits as Hindi number words (e.g. phone numbers are
expanded digit-by-digit: 7155551234 → "सात एक पाँच पाँच पाँच पाँच एक दो तीन
चार"). All phone numbers in this dataset are scripted placeholders (e.g. the
common 555- and 9876543210 dummy patterns), not real numbers. Quantities
(e.g. "30 दिन") are written as digits in the transcript and were converted to
Hindi cardinals via num2words at training/inference time.
Text
- Script is not transliterated or normalized; punctuation including danda (।) is preserved as spoken.
- Code-mixing is preserved as-is (English loanwords/terms like "warranty", "appointment", "repair" appear in Latin script inline with Devanagari).
Files
data/train-*.parquet(12 shards),data/validation-*.parquet(1 shard) — each row has anaudiocolumn (embedded bytes + original filename),text, andspeaker_name. Shards are ~230MB each.
Loading
from datasets import load_dataset
ds = load_dataset("routsourav1729/synthetic-tts")
print(ds)
Known limitations
- 197 segments exceed XTTS's 150-character
hilimit (max observed 182 chars) and may truncate at inference; left in deliberately. - This is a synthetic/scripted speech dataset intended for TTS voice cloning experiments — not a general-purpose ASR corpus.
- Downloads last month
- 84