Rumik OSS 1 releases open weights for expressive Indic text-to-speech
Open Indic voice models are moving toward real model cards, weights, eval protocols, and benchmark artifacts, not only product demos.
Links: Original source · Shared link · Related link 1 · Related link 2 · Related link 3 · Related link 4 · Related link 5
Logged at IST: 2026-09-08 19:46 IST
What it is: Rumik AI's launch of rumik-oss-1, a 3B multilingual text-to-speech model with open weights, focused on expressive Indic and code-switched speech.
Gist: Rumik describes rumik-oss-1 as a 24 kHz text-to-speech model trained on fewer than 70,000 hours of speech. It supports 22 languages, four released voices, native-script and code-switched synthesis, description-conditioned delivery, and inline vocalization tags such as <laugh>, <chuckle>, and <sigh>.
The model card says rumik-oss-1 extends tiny aya fire with discrete speech tokens from the Mimi codec. Text conditioning and audio generation share a single autoregressive sequence; the model predicts eight codec tokens per audio frame and the frozen Mimi decoder reconstructs waveform audio. Speaker identity is separate from delivery controls: prompts can combine tone, accent, and pace while choosing one of the released voices.
Benchmarks: Rumik introduces IndicEmo for expressive delivery in code-switched speech, with prompts spanning English, Hindi, Telugu, Tamil, Kannada, Bengali, and Punjabi. It reports rumik-oss-1 at 2.92/5 overall and 3.03/5 on emotion categories. NoVA evaluates inline laughter, chuckle, and sigh placement; Rumik reports a 0.884 rendering score. The release also reports WER/CER across selected Indic languages using ASR-based transcription comparisons.
Caveats: This is not an unrestricted commercial release. The Hugging Face card lists CC-BY-NC-4.0, inherited from the tiny-aya-derived weights, with an acceptable-use addendum; commercial products and paid synthesis services require separate permission. The technical report, training recipe, and training code are described as forthcoming, and the model card calls out limitations around long-form synthesis and vocalization precision.
Newsletter angle: Interesting because it makes an Indic voice-model release inspectable: weights, demos, samples, benchmarks, eval repos, and a fairly detailed training write-up are all public. The caveat is that "open source" here really means research/non-commercial open weights today, with the most important reproducibility artifacts still promised rather than released.