Datasets:

AkCodes23
/

sarvam-tts-in-te-en

Tasks:

Modalities:

Formats:

Languages:

Size:

Tags:

indian-languages

Libraries:

License:

Dataset card Data Studio Files Files and versions

AkCodes23 commited on 7 days ago

Commit

283ecd3

·

verified ·

1 Parent(s): 7e96981

Upload README.md with huggingface_hub

Files changed (1) hide show

README.md +1 -1

README.md CHANGED Viewed

@@ -55,7 +55,7 @@ Total: **60.06 minutes**.
 - **Single-speaker check** (ECAPA-TDNN embeddings): same-speaker cosine 0.74 vs different-speaker 0.21 (separation 0.52, verification AUC 0.96 / EER 9.1%; 0/11 speakers flagged).
 - **Transcript reliability**: English cross-ASR agreement with Whisper = 6.8% WER / 4.5% CER (n=40) — strong. Realtime ASR language-ID matched the target language on 100% of EN and 100% of TE clips. Telugu cross-ASR is not a valid proxy (Whisper is weak in Telugu); Telugu transcripts are best audited by human review.
 - **Emotion-tag reliability** (sarvam-30b vs sarvam-105b on 120 clips): 65% agreement, Cohen's κ 0.55.
-- **Phoneme coverage**: English 39 (100%), Telugu 45 (90%).
 - **Perceptual quality** (DNSMOS OVRL, published set): EN 3.21 (86% pass>3.0), TE 3.19 (86% pass>3.0). Filter `dnsmos_pass=True` for a stricter subset.
 - **Transcript–audio alignment** (MMS forced-align): median confidence EN 0.948, TE 0.934.

 - **Single-speaker check** (ECAPA-TDNN embeddings): same-speaker cosine 0.74 vs different-speaker 0.21 (separation 0.52, verification AUC 0.96 / EER 9.1%; 0/11 speakers flagged).
 - **Transcript reliability**: English cross-ASR agreement with Whisper = 6.8% WER / 4.5% CER (n=40) — strong. Realtime ASR language-ID matched the target language on 100% of EN and 100% of TE clips. Telugu cross-ASR is not a valid proxy (Whisper is weak in Telugu); Telugu transcripts are best audited by human review.
 - **Emotion-tag reliability** (sarvam-30b vs sarvam-105b on 120 clips): 65% agreement, Cohen's κ 0.55.
+- **Phoneme coverage**: English 39 (100%), Telugu 44 (88%).
 - **Perceptual quality** (DNSMOS OVRL, published set): EN 3.21 (86% pass>3.0), TE 3.19 (86% pass>3.0). Filter `dnsmos_pass=True` for a stricter subset.
 - **Transcript–audio alignment** (MMS forced-align): median confidence EN 0.948, TE 0.934.