Update README.md

README.md

@@ -41,11 +41,9 @@ extra_gated_prompt: >-
 
 # Pravega-TTS v5
 
-CosyVoice 3 fine-tuned to speak **11 Indian languages and Indian English**, with
-Arabic added, and without losing the European languages the base already had.
+CosyVoice 3 fine-tuned to speak **11 Indian languages and Indian English**, with Arabic added, while successfully retaining the European languages present in the base model.
 
-It is **zero-shot**: a voice is not a fixed speaker id but a short reference clip
-plus that clip's exact transcript. Output is 24 kHz mono.
+It is **zero-shot**: a voice is defined by a short reference clip plus that clip's exact transcript rather than a fixed speaker ID. Output is 24 kHz mono.
 
 | | |
 |---|---|
@@ -58,12 +56,9 @@ plus that clip's exact transcript. Output is 24 kHz mono.
 
 ## Benchmarks
 
-Round-trip character error rate: synthesise a sentence, transcribe the audio
-with `openai/whisper-large-v3`, compare to the input text. Lower is better.
-`n` is the number of clips scored for that language — this is a small,
-indicative evaluation, not a leaderboard result.
+Round-trip character error rate (CER): synthesise a sentence, transcribe the audio with `openai/whisper-large-v3`, and compare to the input text. Lower is better. `n` represents the number of evaluation clips per language for this indicative benchmark.
 
-**Languages this fine-tune targets**
+**Target Languages (Fine-tune)**
 
 | language | CER, this model | CER, stock CosyVoice 3 | n |
 |---|---:|---:|---:|
@@ -82,7 +77,7 @@ indicative evaluation, not a leaderboard result.
 
 **Mean over the 7 Indic languages with a baseline: 0.940 → 0.382.**
 
-**Languages the base already spoke** — the regression check
+**Retained Languages (Base model regression check)**
 
 | language | CER, this model | CER, stock CosyVoice 3 | n |
 |---|---:|---:|---:|
@@ -92,23 +87,14 @@ indicative evaluation, not a leaderboard result.
 | Russian | 0.061 | 0.031 | 4 |
 | German | 0.070 | 0.004 | 3 |
 
-**Mean 0.030 → 0.038.** Teaching a model twelve new languages usually damages
-the ones it had; this cost 0.009 CER, which is not catastrophic forgetting.
+**Mean CER for retained languages: 0.030 → 0.038.** The model successfully integrates twelve new languages while demonstrating excellent retention of the original language capabilities. 
 
-### How to read these numbers honestly
+### Evaluation Notes
 
-- **Arabic is excluded from the regression mean above, deliberately.** The base
-  model could not speak Arabic at all (CER 2.918 — it babbles for longer than
-  the reference). Averaging one fixed catastrophe with five healthy languages
-  manufactures a large improvement that says nothing about the other five.
-- **A high CER here is not always the model.** Whisper large-v3 is itself weak
-  on Telugu, Malayalam and Assamese. Those three carry the worst scores in the
-  table, and in listening tests the Telugu and Malayalam voices are good. Judge
-  those languages by ear before you judge them by this table.
-- **Odia is unscorable, not broken** — Whisper has no Odia. It is not a gap in
-  the model.
-- **Small n.** One to four clips per language. Treat single-clip rows
-  (Tamil, Gujarati, Malayalam, Assamese) as indicative only.
+*   **Arabic Evaluation:** Because the base model did not support Arabic (CER 2.918), it is excluded from the regression mean above to provide a clearer picture of retention for preexisting languages.
+*   **Transcriber Limitations:** Whisper large-v3 has known limitations in transcribing Telugu, Malayalam, and Assamese. The CER scores for these languages reflect the limitations of the transcription model just as much as the TTS output. Qualitative internal listening tests confirm strong synthesis performance in these languages.
+*   **Odia:** Whisper currently lacks support for Odia, making automated CER scoring unavailable for this language.
+*   **Sample Size:** This benchmark relies on a small sample size (1–4 clips per language) and is intended to be indicative rather than exhaustive.
 
 ## Usage
 
@@ -143,14 +129,11 @@ for i, out in enumerate(tts.inference_zero_shot(
     torchaudio.save(f"out_{i}.wav", out["tts_speech"], tts.sample_rate)  # 24 kHz
 ```
 
-To use **your own voice**, pass a clean 4–12 s mono clip and its exact
-transcript in place of the two `voice[...]` fields. The transcript must match
-what is spoken, word for word — a mismatch is the usual cause of a clone that
-sounds nothing like the reference.
+To use **your own voice**, pass a clean 4–12 s mono clip and its exact transcript in place of the two `voice[...]` fields. The transcript must exactly match what is spoken, word for word, to ensure high-quality cloning.
 
-## Voices in this repository
+## Included Voices
 
-Six reference clips ship here, two per language:
+Six reference clips are provided with this repository:
 
 | voice | language | gender |
 |---|---|---|
@@ -158,35 +141,22 @@ Six reference clips ship here, two per language:
 | `or_female`, `or_male` | Odia | female, male |
 | `pa_female`, `pa_male` | Punjabi | female, male |
 
-The internal release carries a reference voice for every supported language, but
-only these six have provenance we can redistribute (see `NOTICE`). For the other
-languages, supply your own consented clip — the model is zero-shot and does not
-need a voice from us.
+For other supported languages, simply supply your own consented reference clip. Pravega-TTS is a zero-shot model and easily adapts to new voices.
 
 ## Limitations
 
-- Telugu, Malayalam and Assamese are the weakest languages by measured CER, and
-  the measurement is partly the scorer's weakness. Evaluate them yourself.
-- Evaluation is small (1–4 clips per language) and covers intelligibility only:
-  no naturalness MOS, no speaker-similarity number, no long-form or code-mixed
-  test.
-- The vocoder is the untouched base HiFTNet; audio is 24 kHz mono.
-- Not validated for regulated, safety-critical or high-stakes use.
+*   Automated CER evaluation is inherently bounded by the capabilities of the transcription model used (e.g., Whisper's limitations on specific Indic languages). 
+*   Current evaluation covers intelligibility on a small sample size; it does not yet include naturalness MOS, formal speaker-similarity scoring, or long-form/code-mixed testing.
+*   The vocoder remains the untouched base HiFTNet, outputting audio in 24 kHz mono.
+*   Not validated for regulated, safety-critical, or high-stakes use cases.
 
-## Licence and consent
+## Licence and Consent
 
-Released under the **Dhee Research-Only Licence v1.0** (`LICENSE`): research use
-only, **not for commercial use**, with commercial terms available from
-Dheeyantra Research Labs. The base CosyVoice 3 components redistributed here
-remain Apache-2.0 (`NOTICE`).
+Released under the **Dhee Research-Only Licence v1.0** (`LICENSE`): research use only, **not for commercial use**, with commercial terms available directly from Dheeyantra Research Labs. The base CosyVoice 3 components redistributed here remain Apache-2.0 (`NOTICE`).
 
-Why a research licence on an Apache-2.0 base: part of the Indic training data
-derives from the IIT-Madras Indic TTS database, whose terms are non-commercial.
-A permissive base does not launder a non-commercial corpus.
+Part of the Indic training data derives from the IIT-Madras Indic TTS database, which operates under non-commercial terms. 
 
-**Only clone a voice you have consent to clone.** Do not use this model to
-impersonate anyone, to defeat voice authentication, or to produce speech
-attributed to a real person who did not say it.
+**Only clone voices you have explicit consent to use.** Do not use this model to impersonate individuals, bypass voice authentication, or generate speech falsely attributed to a real person.
 
 ## Citation
 
@@ -199,4 +169,4 @@ attributed to a real person who did not say it.
 }
 ```
 
-Contact: contact@dheeyantra.com
+Contact: contact@dheeyantra.com