Prakhar Khatri

Tinkering with voice...

[Prakhar Khatri @PrakharKhatri3]

Introducing IndiCodec: a 1.09 kbps neural speech codec for speech LMs and TTS, one of a kind, built for english and indian languages.

Strongest on Indian languages: On 10 Indian languages, it adds 9x less error than Mimi, with 25% fewer tokens.

In English, it has the lowest WER of the ~1 kbps codecs we tested.

Apache-2.0. Open source today!

[Prakhar Khatri @PrakharKhatri3]

the architecture of Indicodec:)

[Prakhar Khatri @PrakharKhatri3]

@Dheemanthredy @Alibaba_Qwen @kyutai_labs congrats, this is great to see!
we also need better audio codecs trained on indian languages, built for Indic speech LMs:)
starting here: https://t.co/pMpIDHr9mn

[Prakhar Khatri @PrakharKhatri3]

- It is based on a split FSQ architecture(unlike Neucodec/SNAC) with two separate streams for content(what is spoken) and acoustic(how it is spoken, residual of content) streams, both coded separately. Comprising two FSQ grids.
- No RVQ stack(unlike Mimi): one FSQ token per stream per frame, so the LM has fewer tokens to model
- Built on Neucodec, specifically trained on indian languages for better Indic performance

[Prakhar Khatri @PrakharKhatri3]

Introducing IndiCodec: a 1.09 kbps neural speech codec for speech LMs and TTS, one of a kind, built for english and indian languages.

Strongest on Indian languages: On 10 Indian languages, it adds 9x less error than Mimi, with 25% fewer tokens.

In English, it has the lowest WER of the ~1 kbps codecs we tested.

Apache-2.0. Open source today!

[Prakhar Khatri @PrakharKhatri3]

@AcousimHss try indicodec once:)

[Prakhar Khatri @PrakharKhatri3]

Introducing IndiCodec: a 1.09 kbps neural speech codec for speech LMs and TTS, one of a kind, built for english and indian languages.

Strongest on Indian languages: On 10 Indian languages, it adds 9x less error than Mimi, with 25% fewer tokens.

In English, it has the lowest WER of the ~1 kbps codecs we tested.

Apache-2.0. Open source today!

[shobhit @Sinister_Light]

Prakhar is solving a foundational component for Voice models - super proud

[Prakhar Khatri @PrakharKhatri3]

@neuphonicspeech thanks team
love the work Neuphonic is doing!

[Prakhar Khatri @PrakharKhatri3]

@Sinister_Light Thanks shobhit, means a lot!

[Neuphonic @neuphonicspeech]

Love the work! Nicely built upon our open-source codec.

[Prakhar Khatri @PrakharKhatri3]

Links:

- Model weights (HF): https://t.co/pMpIDHr9mn
- Code (GitHub): https://t.co/8ahhKkY24Q
- Demo page with audio samples and results: https://t.co/Tb8YMqZsf1

[Prakhar Khatri @PrakharKhatri3]

Built on NeuCodec by @neuphonicspeech , with a new two-stream quantizer: 75 tokens/s, 25% fewer than Mimi. One stream carries the words (25 Hz), the other the voice (50 Hz).
No RVQ stack: one token per stream per frame.

[Prakhar Khatri @PrakharKhatri3]

tommorow is the day!!

[Prakhar Khatri @PrakharKhatri3]

big week ahead🤞

[Prakhar Khatri @PrakharKhatri3]

there is so much to build!

[jack @jack]

We believe voice is the most natural way to interact with intelligence, so we created Dessa, the most authentic full-duplex speech model ever built. Talking with it feels like talking with a person…full of emotion, nuance, pauses, and a natural cadence.

https://t.co/r0qFLdvd6c

[Prakhar Khatri @PrakharKhatri3]

@tarini_axory @DetectIfAI congrats on the launch!!

[Prakhar Khatri @PrakharKhatri3]

big week ahead🤞

[Elon Musk @elonmusk]

Turns out the future does have flying cars

[Prakhar Khatri @PrakharKhatri3]

@FishAudio Drama