Rumik-OSS-1: The First Open-Source Indic TTS With Emotion

Open-source text-to-speech has spent the last two years catching up to closed voice APIs on English. Indic languages have mostly waited. On September 8, 2026, Bengaluru-based Rumik Intelligence released Rumik-OSS-1, a 3-billion parameter open-weight model it calls "the first open source indic text to speech model with expressive emotions" — spanning 22 languages, code-switched speech, and inline emotion control, with weights already live on Hugging Face.

That claim is specific enough to check. Most Indic TTS work to date — including Sarvam's Bulbul and BharatGen out of IIT Bombay — has shipped as closed or semi-closed APIs. Rumik-OSS-1 is the first entry in that category to publish weights, even if the license keeps it out of production for now.

TL;DR — what people are asking

What Rumik-OSS-1 actually ships

Strip away the framing and four things are confirmed as of launch:

1. A 3B-parameter model fine-tuned from Cohere's Tiny Aya Fire, released in BF16 on Hugging Face.
2. 22 Indic languages plus English, in both native script and romanized transcription, with support for code-switched sentences that mix languages mid-utterance — a real requirement for Hinglish and similar everyday speech patterns.
3. Parametric and inline control: pace, accent and tone as parameters, plus inline tags for laughter, sighs and chuckles placed at specific points in the text, alongside natural-language style descriptions.
4. Four named voices — Ira, Aisha, Siya and Zoya — rather than open-ended voice cloning from a reference clip.

That last point matters for how the model should be read. Rumik-OSS-1 is not a clone-from-five-seconds system in the mold of Fish Audio's S2 family. It's closer in philosophy to Sarvam's curated-voice approach: a fixed voice library plus expressive range, which reduces impersonation risk but also means you can't hand it a reference clip and get a new speaker.

The architecture: Tiny Aya Fire meets the Mimi codec

Rumik-OSS-1's technical shape is a decoder-only transformer doing next-token prediction over audio tokens, not a diffusion or flow-matching pipeline. Two components carry the weight:

The model predicts eight codebook tokens per audio frame before advancing to the next frame — a standard residual-vector-quantization approach to neural audio codecs, similar in spirit to the tokenization strategy behind models like VoxCPM2, though Rumik's approach is token-based rather than tokenizer-free.

Training pipeline

Rumik's model card and research page describe three stages:

1. Pre-training (75,000 updates) — an English speech foundation using the Emilia dataset, roughly 40,000 hours.
2. Multilingual expansion (30,000 updates) — Indian-language data from IndianVoices, IndianVoices-R and RASA, roughly 10,200 hours combined.
3. Post-training with GRPO — Group Relative Policy Optimization, the same reinforcement-learning technique behind several recent open-weight reasoning and coding releases, applied here to optimize for reward signals tied to expressive delivery rather than task correctness.

Under 70,000 total hours of speech data is notably lean next to some proprietary TTS training runs, and Rumik's own materials lean on that as a selling point — the model "performs competitively... while trained on fewer than 70,000 hours."

Which languages does it actually support?

Rumik's research page and model card both state 22 Indic languages, in native script and romanized form, plus English — but neither publishes the full 22-language list as a table. The languages explicitly named across Rumik's materials are:

• Hindi
• Bengali
• Tamil
• Telugu
• Kannada
• Punjabi
• English

That's 7 confirmed by name; the remaining ~15 aren't individually enumerated in the public research page or model card. Given the "22" figure and Rumik's Indic framing, this almost certainly maps to (or closely tracks) India's Eighth Schedule list of 22 scheduled languages — the same reference list BharatGen uses — but Rumik hasn't confirmed that mapping in writing, so treat it as a strong inference, not a verified fact. If you need a specific language (Odia, Malayalam, Gujarati, Marathi, Assamese, Urdu, etc.), the only way to confirm coverage today is to test it directly against your target language.

How to use Rumik-OSS-1

Online (hosted demo)

Rumik has a Hugging Face Space set up for browser-based testing with no local install. At the time of writing, the Space was still initializing ("Fetching metadata..." / "Refreshing") and not reliably serving requests — it runs on Hugging Face's shared "Zero" GPU tier, so availability may be intermittent depending on demand. Check it directly before assuming it's live; there is no other official hosted playground or API endpoint published yet.

Locally (self-hosted)

Running it yourself is the more dependable path right now. Rumik ships a server.py that exposes an OpenAI-compatible speech endpoint. The documented setup:

# 1. Set up a Python 3.12 virtual environment
python3.12 -m venv .venv
source .venv/bin/activate

# 2. Install the Hugging Face CLI and download the weights
python -m pip install "huggingface_hub==0.36.2"
hf download rumik-ai/rumik-oss-1 --local-dir ./rumik-model

# 3. Install the model's own dependencies
python -m pip install -r ./rumik-model/requirements.txt

# 4. Launch the local web server
python ./rumik-model/server.py --repo-id ./rumik-model --host 127.0.0.1 --port 6006

Once running, the server accepts requests against a /v1/audio/speech endpoint (an OpenAI TTS-API-shaped route), taking speaker, input text and generation parameters in the request body — familiar to anyone who has already integrated an OpenAI-compatible TTS client.

Hardware: the model card specifies an NVIDIA GPU with CUDA support as a requirement; no CPU inference path or specific VRAM figure is published. At 3B parameters in BF16, expect to need a consumer GPU with at least 8–12GB VRAM as a practical floor, though Rumik hasn't confirmed an exact number — benchmark your own hardware before assuming it fits.

What's next for Rumik-OSS-1

Rumik's research page lists several things explicitly as coming, not yet shipped:

• A full technical report, detailing the training methodology in more depth than the current research page.
• Training code, released alongside the technical report — Rumik's own phrasing is "yes you heard that right," implying this wasn't a given.
• Upgraded checkpoints with more consistent emotion-description handling — an acknowledgment that the current model's emotion tagging isn't fully stable yet.
• Further post-training work on decoder-only transformer speech models generally — phrased as an open-ended research direction rather than a specific dated release.

Notably absent from the public roadmap: a commercial license path, a hosted paid API, voice-cloning support, or a firm timeline for any of the above. If your use case depends on one of those, the only way to get a straight answer right now is to email research@rumik.ai directly — Rumik has invited exactly that for implementation questions.

The benchmarks, and their limits

Rumik reports two self-designed evaluation metrics:

Two things are worth flagging plainly. First, both indicEmo and NoVA are Rumik's own benchmarks, evaluated across 98–149 prompts per test — useful as a self-report, but not yet cross-checked by an independent leaderboard the way Cartesia's Sonic-3.6 has been on Artificial Analysis. Second, Rumik's own model card discloses a real limitation: NoVA detectors report unrequested vocalizations in 2.9–5.4% of outputs — meaning the model sometimes inserts a laugh or sigh nobody asked for. That's the kind of number a self-published benchmark rarely volunteers, and it's a point in favor of Rumik's transparency even as it flags a genuine reliability gap.

The model is also trained and validated on utterances up to 30 seconds, with Rumik cautioning against pushing past 35 seconds — a real constraint for anything beyond short-form prompts, IVR responses or dialogue lines.

Benchmark table: Rumik-OSS-1 vs. the field

Here's the honest version of this table, not the marketing version. Rumik has not published a head-to-head run against any named competitor — indicEmo and NoVA are Rumik's own benchmarks, and no other TTS vendor has reported scores on them. Anyone showing you a table with Sarvam or ElevenLabs slotted into an indicEmo column is inventing numbers. What follows is what's actually documented for each model, on whatever metric that vendor chose to publish:

The takeaway from this table isn't "Rumik wins" or "Rumik loses" — it's that Rumik-OSS-1 is the only model in this set that is both open-weight and benchmarked on expressiveness specifically, and it's also the only one grading itself on its own curve. Cartesia is the one model here with genuine third-party leaderboard placement; everyone else, Rumik included, is self-reported until an outside evaluator runs the same prompts across all of them.

What's missing at launch

Rumik's research page is explicit that some pieces are still coming:

• Training code — not yet released, so reproducing the pipeline isn't possible today.
• Technical report — listed as forthcoming; the research page and model card are the primary documentation for now.
• Independent benchmark verification — no third-party leaderboard placement yet.
• Commercial license path — the model card points to contacting Rumik directly rather than a published pricing or licensing page.

None of that makes the release less real — weights are live, and that's the part builders can act on today. But it does mean Rumik-OSS-1 should be evaluated as a research/non-commercial release right now, not treated as production-ready infrastructure.

How it fits next to Sarvam, BharatGen and the rest

India's Indic-language AI landscape has been building for a while — India's sovereign AI push under the IndiaAI Mission has funded model development across several labs, and Sarvam's full API stack already covers speech recognition, translation and TTS as hosted products. What's different about Rumik-OSS-1 is the axis it competes on: openness, not polish.

None of these are apples-to-apples yet — Sarvam's Bulbul V4 launched with a demo reel and no public model ID, and BharatGen is positioned more as sovereign-AI infrastructure than a drop-in TTS API. What Rumik-OSS-1 adds to that landscape is a checkpoint anyone can download and test today, on any of 22 languages, without an API key or waitlist — a genuinely different distribution model even before the license question is settled.

Should you use it?

For research, prototyping, or evaluating whether expressive Indic TTS is viable for your use case: yes, worth downloading today. For production: not yet, on two counts — the non-commercial license blocks it outright, and the unrequested-vocalization rate (2.9–5.4%) means transactional or safety-critical speech (OTPs, account balances, legal disclosures) needs guardrails before this model touches them.

A reasonable evaluation plan mirrors what explainx.ai has recommended for other emerging TTS launches: build a script corpus spanning transactional, support, education and code-mixed Indian-language sentences, render with Rumik-OSS-1 and a documented baseline, and score on word accuracy, pronunciation of names, style-tag adherence, and — specifically for this model — false-positive vocalization rate.

Honest limitations

• Both published benchmarks (indicEmo, NoVA) are self-reported, with no independent leaderboard placement yet.
• Training code and the full technical report are not yet released.
• License is non-commercial (CC-BY-NC 4.0 + acceptable-use addendum); commercial use requires direct permission from Rumik.
• The model reports unrequested vocalizations in up to 5.4% of outputs per its own card.
• Training and evaluation are capped around 30–35 second utterances.
• As of documentation, Hugging Face listed only 12 downloads in the prior month — this is a fresh release, not yet battle-tested by the community.

Final verdict

Rumik-OSS-1 is a genuine first, not a finished product. It is the first open-weight Indic TTS model with expressive emotion control that anyone can download today — that specific claim holds up. It is not the best-sounding Indic voice model available (Sarvam and Cartesia both have more production mileage), it is not independently benchmarked, and it is not usable in a commercial product without contacting Rumik first.

Score it like this:

If you're building a research project, a language-preservation tool, or a proof-of-concept for expressive Indic speech, download it today. If you're shipping a commercial product, treat this as a preview of where open Indic TTS is headed in the next 6–12 months, not the model you ship with.

Does this mean India is winning the AI race?

Not by itself — but it's a real data point in a broader pattern worth naming honestly. Three things are true at once:

What's genuinely encouraging: India now has multiple independent efforts attacking Indic-language AI from different angles in the same year — Sarvam's hosted commercial stack, IIT Bombay's BharatGen as government-backed sovereign infrastructure, and now Rumik as an open-weight research release. That's a healthier ecosystem than one company or one lab carrying the language-coverage problem alone, and it mirrors the diversification the IndiaAI Mission has been explicitly funding.

What's still missing: none of these are frontier-model efforts in the sense that DeepSeek, Kimi or the major US/China labs are frontier-model efforts. Indic TTS at 3B parameters, trained on under 70,000 hours, is an important and underserved problem — but it's a specialized, regionally-scoped model, not evidence of India competing on foundation-model scale, compute, or general-purpose capability. The AI race that dominates headlines (frontier LLMs, trillion-parameter training runs, global model rankings) is a different competition than the one Rumik just entered.

What actually matters here: language coverage is not a side quest. Over 600 million people in India don't have production-quality, expressive AI speech in their first language today, in any product, from anyone. A 3B open-weight model that starts closing that gap — even imperfectly, even non-commercially — moves a real problem forward in a way a bigger, better-funded frontier lab has no particular incentive to prioritize. India isn't winning "the AI race" as a monolith. It is building genuine depth in a dimension — linguistic diversity at scale — that the leading labs mostly ignore, and that's a legitimate, separate kind of progress worth tracking on its own terms rather than folding into a single global leaderboard.

The explainx.ai read

The "first open-source" claim is defensible on the specific axis Rumik is claiming — expressive, emotion-controllable Indic TTS with published weights. It's a narrower claim than "best Indic TTS," and Rumik's own materials mostly stay inside that boundary rather than overselling it. The more interesting signal is what it says about where Indic-language voice AI is heading: after a year of hosted APIs from Sarvam and institutional efforts like BharatGen, an open-weight checkpoint anyone can download changes who gets to build on top of it, even under a restrictive license. The commercial story and the independent verification are still to come — track the technical report and any leaderboard placement before treating the benchmark numbers as settled.

Related on explainx.ai

• Sarvam Bulbul V4: more expressive Indian voice model — the closest hosted competitor, launched five weeks earlier
• BharatGen: IIT Bombay's sovereign AI for all 22 scheduled languages
• Sarvam AI: full capabilities guide
• India's sovereign AI status under the IndiaAI Mission
• Cartesia Sonic-3.6: #1 on Artificial Analysis TTS boards
• VoxCPM2: tokenizer-free multilingual voice cloning
• Intelligence Ownership: GRPO fine-tuning explained
• Why explainx.ai supports open-source AI

Primary sources

• Rumik-OSS-1 research overview
• Rumik-OSS-1 model card on Hugging Face

Model specifications, benchmark scores and license terms are accurate as of September 8, 2026, the publication date. Rumik has indicated training code and a full technical report are forthcoming — check the primary sources above before building on details that may have since been published.