rumik-oss 1¶

rumik-oss 1 is a 3B expressive, multilingual text-to-speech model from rumik.ai for 22 Indic languages plus English, in native scripts and romanized forms, with code-switched synthesis. A decoder-only language model emits eight Mimi codec tokens per 80 ms frame and the frozen codec reconstructs 24 kHz audio. Delivery is controlled with a text description (tone, accent, pace) and inline vocalization tags.

The MLX port loads the original Hugging Face repo directly, applies the model's stop predictor, restricts sampling to the audio vocabulary, and streams audio through mlx-audio's native Mimi decoder.

Model Variants¶

Model	Format	HuggingFace
`rumik-ai/rumik-oss-1`	BF16 safetensors (original repo)	Model Card
`rumik-ai/rumik-oss-1-mlx-8bit`	8-bit MLX quantized (recommended)	Model Card
`rumik-ai/rumik-oss-1-mlx-4bit`	4-bit MLX quantized (fastest, lower fidelity)	Model Card

The Mimi codec weights are fetched from kyutai/moshiko-pytorch-bf16 on first use (the codec bundled with the original repo is byte-identical). Requires ~7 GB of memory in BF16, ~4 GB at 8-bit, ~2 GB at 4-bit.

Usage¶

Basic Generation¶

CLIPython

mlx_audio.tts.generate \
    --model rumik-ai/rumik-oss-1-mlx-8bit \
    --text "नमस्ते, आज आपका दिन कैसा रहा?" \
    --voice Ira \
    --instruct "happy, Hindi accent, steady pace"

from mlx_audio.tts.utils import load_model

model = load_model("rumik-ai/rumik-oss-1-mlx-8bit")

for result in model.generate(
    text="नमस्ते, आज आपका दिन कैसा रहा?",
    voice="Ira",
    instruct="happy, Hindi accent, steady pace",
):
    audio = result.audio  # mx.array, 24 kHz

Streaming¶

Audio is yielded in chunks as frames are generated; streaming_interval sets the chunk length in seconds.

CLIPython

mlx_audio.tts.generate \
    --model rumik-ai/rumik-oss-1-mlx-8bit \
    --text "Good evening, passengers. Boarding will begin shortly." \
    --voice Aisha \
    --instruct "professional, Indian English accent, steady pace" \
    --stream --streaming_interval 0.5 --play

from mlx_audio.tts.utils import load_model

model = load_model("rumik-ai/rumik-oss-1-mlx-8bit")

for chunk in model.generate(
    text="Good evening, passengers. Boarding will begin shortly.",
    voice="Aisha",
    instruct="professional, Indian English accent, steady pace",
    stream=True,
    streaming_interval=0.5,
):
    play(chunk.audio)  # chunk.is_final_chunk marks the last one

Delivery Controls¶

The instruct argument becomes the <description="..."> prefix of the prompt. The description can also be written inline in the text, and vocalization tags are placed where they should occur.

model.generate(
    text='<description="excited, Hindi accent, fast pace"> जल्दी आओ, हमारा नाम लिस्ट में है! <laugh> आज घर में जश्न होगा।',
    voice="Siya",
)

Control	Values
`voice`	`Ira`, `Aisha`, `Siya`, `Zoya`
tone	happy, sad, angry, excited, professional
accent	Hindi, Telugu, Tamil, Kannada, Bengali, Punjabi, Indian English
pace	slow, fast, steady
inline tags	`<laugh>`, `<chuckle>`, `<sigh>`

Generation Parameters¶

Parameter	Default	Description
`temperature`	0.8	Sampling temperature (reference setting)
`top_k`	30	Top-k sampling (reference setting)
`max_tokens`	2048	Cap on audio tokens; 100 tokens make one second of speech
`min_tokens`	8	Tokens generated before the utterance may end
`streaming_interval`	0.5	Seconds of audio per streamed chunk

Generation ends when the model's stop predictor fires or </audio> is sampled, matching the reference implementation. Utterances longer than about 35 seconds are outside the model's training distribution.

Performance¶

Decode speed on an M5 MacBook Air (16 GB), steady state, audio head sliced to the audio vocabulary:

Variant	Tokens/s	Real-time factor	Peak memory	Next-token agreement vs. BF16
BF16	21	0.21x	6.9 GB	--
8-bit	33-40	0.33-0.40x	3.8 GB	96%
4-bit	69	0.69x	2.1 GB	80%

The flattened codec formulation runs the 3B backbone once per codebook token, so 100 forward passes produce one second of audio. Streaming delivers first audio in well under a second, but sustained real-time playback on Apple silicon laptops is not yet reached.

License¶

rumik-oss 1 weights are released for research and non-commercial use under CC-BY-NC 4.0 with the Cohere Labs acceptable-use addendum. The Mimi codec is CC-BY-4.0.