Swarit Indic Talk
Streaming voice-to-voice conversations across Indic languages.
Swarit Indic Talk makes voice the natural way to have a conversation. Speak in your language, let Swarit follow what you mean, and hear the reply spoken back. Built for conversations across India—in homes, clinics, shops, classrooms, and customer support.
Visit Swarit.ai · Download Swarit Indic Talk
Voice-to-voice in 12 languages · Cross-language voice cloning · Low-latency streaming · Conversational memory · Natural turn-taking
Watch and listen
Hear Swarit carry an appointment conversation from the first question to the next step.
One voice. Twelve languages.
Speaker-conditioned voice cloning carries a chosen voice across supported languages. Swarit listens, follows the conversation, and speaks back with low-latency streaming, conversational memory, and natural turn-taking.
A voice-to-voice experience across 12 languages
Swarit Indic Talk brings spoken conversation to eleven Indian languages—Assamese, Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, and Telugu—and English. People can begin in the language they are comfortable speaking and hear the conversation continue in voice.
Language communities, one conversational model
- East and Northeast: Assamese, Bengali, and Odia bring voice conversations to communities across Assam, West Bengal, Tripura, and Odisha.
- North: Hindi and Punjabi support spoken exchanges across Hindi- and Punjabi-speaking communities.
- West: Gujarati and Marathi extend voice conversations across Gujarat and Maharashtra.
- South: Kannada, Malayalam, Tamil, and Telugu bring the experience to communities across Karnataka, Kerala, Tamil Nadu, Andhra Pradesh, and Telangana.
- English: English is included for multilingual conversations and for people who prefer to speak in English.
The language names below are shown in their native forms, alongside the writing systems associated with them.
| Language | Native name | Writing system |
|---|---|---|
| Assamese | অসমীয়া | Assamese |
| Bengali | বাংলা | Bengali |
| English | English | Latin |
| Gujarati | ગુજરાતી | Gujarati |
| Hindi | हिन्दी | Devanagari |
| Kannada | ಕನ್ನಡ | Kannada |
| Malayalam | മലയാളം | Malayalam |
| Marathi | मराठी | Devanagari |
| Odia | ଓଡ଼ିଆ | Odia |
| Punjabi | ਪੰਜਾਬੀ | Gurmukhi |
| Tamil | தமிழ் | Tamil |
| Telugu | తెలుగు | Telugu |
Voice in. Conversation out.
Swarit brings the full conversational experience together: it listens to spoken input, follows the exchange, and answers in speech. The result is voice-to-voice interaction that feels direct and familiar—without making people adapt to a text-first interface.
Use Swarit for the conversations that move everyday life forward: booking an appointment, asking about an order, getting help from a business, or finding the next step in an application.
Release scope and architecture
The downloadable v1.0.0 artifact is a weights-only bundle for a modular speech-to-speech pipeline. It combines separately trained components; it is not a single jointly trained end-to-end audio model. The release selection is quantized-20261016 in manifest.json.
| Component | Published v1 selection |
|---|---|
| ASR | Qwen3-ASR 1.7B, fine-tuned epoch 3 / cumulative step 5391, BF16 |
| Response generation | Gemma 4 E4B, fine-tuned epoch 2; NF4 base linears, INT4 per-layer embeddings and a separate, unmerged LoRA adapter |
| Speech generation | Qwen3-TTS 12Hz 0.6B, fine-tuned epoch 12 / step 26264; BF16 talker/predictor and FP32 codec |
| Conversation detector | wav2vec2 XLS-R 300M, acoustic epoch 9 / policy head 29; speech activity, language, turn completion and conversational policy |
| Conversation memory | XLM-R Large XNLI pinned encoder with adapters and state heads |
| Supplied voices | Two offline-extracted float32 speaker vectors, reused across output languages |
Stored package sizes are not runtime GPU memory requirements. Runtime memory also includes activations, caches and other overhead. The detector's compact packed weights do not imply integer-kernel inference: its learned quantized values are materialized for runtime computation.
Hosted system versus this download (11 November 2026): the hosted system has evolved to a later Qwen3-ASR checkpoint (epoch 26), Sarvam 105B for responses and VoxCPM2 for TTS, with 24 approved speakers and three reference-based delivery modes. These components and controls are not supplied by the v1 download. The evaluation results below name their exact checkpoint and must not be attributed to the newer hosted stack.
Streaming and conversation behavior
Streaming is a core part of the serving runtime: incoming speech is processed incrementally, generated reply text is streamed into TTS, and audio chunks are sent for playback before the complete reply finishes. The runtime supports receiving audio during speech output and barge-in cancellation of an ongoing reply. These behaviors require the controller, transport and playback integration, beyond the weights themselves.
Overlapping backchannels and robust full-duplex interaction across languages and devices have not been comprehensively validated. Implemented behavior is best described as streaming conversation with interruption support. Other ASR systems, including SraVaani 0.5-live, also support streaming; it is not an exclusive capability. Streaming and diffusion describe different aspects of a system and are not mutually exclusive.
Evaluation results and limitations
ASR: five models on identical recordings
Fresh inference ran on one RTX 3090 using the fixed asr-corpus-validation-1k/20261009-v1 internal validation set. The common comparison below contains 916 identical recordings across eleven Indic languages. WER and CER use shared project normalization and standard edit counts; CER counts Unicode code points with spaces removed. Lower is better.
| Model | WER | CER |
|---|---|---|
| IndicConformer 600M, native RNNT | 27.42% | 11.75% |
| Bodhan Indic-Transcribe Flex | 28.60% | 12.23% |
| SraVaani 0.5-live, native streaming CTC | 31.38% | 13.85% |
| Fine-tuned Qwen3-ASR 0.6B, epoch 56 / step 26264 | 31.97% | 13.56% |
| Fine-tuned Qwen3-ASR 1.7B, epoch 3 / step 5391 (v1 ASR) | 32.50% | 15.16% |
The same validation set contains 84 English recordings, scored separately because IndicConformer does not support English:
| Model | WER | CER |
|---|---|---|
| Fine-tuned Qwen3-ASR 1.7B, epoch 3 / step 5391 (v1 ASR) | 3.29% | 1.02% |
| Bodhan Indic-Transcribe Flex | 3.44% | 0.89% |
| Fine-tuned Qwen3-ASR 0.6B, epoch 56 / step 26264 | 8.68% | 3.95% |
| SraVaani 0.5-live | 13.51% | 4.90% |
Qwen, Bodhan and IndicConformer received the registered language; SraVaani used automatic language handling. Native decoding paths and batch sizes differed. Empty and capped outputs remained in scoring. These are internal development recordings with documented historical training exposure and speaker overlap, not an independent held-out benchmark or publisher leaderboard. The tables measure transcription accuracy, not the complete streaming conversation system.
The same v1 ASR checkpoint separately measured 22.66% WER / 5.49% CER on a different 560-clip, twelve-language development set. That result uses different recordings and cannot be directly compared with the five-model tables above. WER is specific to a dataset and protocol, rather than a fixed property of a checkpoint.
TTS intelligibility
The published epoch-12 checkpoint, with BF16 talker/predictor and FP32 codec, measured 18.53% WER / 5.72% CER through Bodhan Indic-Transcribe Core on 500 generated utterances across twelve languages. Prompts were 458 Rasa test captions and 42 LibriTTS-R test.clean English captions, using one fixed Rasa speaker.
Scoring used supplied language, native script and beam 5, with no target-text hints. These are recognizer-based intelligibility indicators, not human naturalness, speaker-similarity or MOS scores. This controlled component evaluation does not evaluate the complete live serving runtime, and historical training exposure is incompletely audited.
Turn completion, MOS and latency
The selected acoustic epoch-9 detector scored 85% turn-completion accuracy on 500 clip-end examples from Pipecat's smart-turn-data-v3.1-train test split. It recorded 60 false completions and 15 missed completions. This is a clip-end classification result, not measured live endpoint delay or interruption quality; speaker-disjointness and broad multilingual coverage are unverified.
No formal human-rated MOS study is reported. Configuration-specific server latency measurements exist, but there is no comprehensive mouth-to-ear p50/p95 benchmark across all twelve languages. Server first-PCM timing, TTS real-time factor and perceived playback latency are distinct measurements. The available results do not establish an overall best-model ranking across Indic languages or complete conversational quality.
Non-verbal cues and expressive output
The wav2vec detector processes audio directly for speech activity, language, turn completion and policy prediction. Response generation primarily consumes transcripts and conversation state. Reliable understanding of laughter, breathing, hesitation and vocal emotion has not been validated.
The published v1 bundle does not expose validated controls for laughter, sighs or explicit emotions. The newer hosted VoxCPM2 system uses approved natural, happy and empathetic reference recordings with manual or automatic text-based selection. That is separate from acoustic emotion recognition and is not included in v1. Broad perceptual and Indic tone-accuracy evaluation remains pending.
Speaker conditioning and voice cloning scope
The two supplied vectors were extracted offline with the original speaker encoder from selected IISc SYSPIN/LIMMITS24 recordings. Inference loads the vectors directly and keeps the chosen speaker fixed across output languages, so a resident speaker encoder is unnecessary.
The download supports conditioning on the two supplied voices. Creating a new vector from an arbitrary recording requires the matching encoder and preprocessing, which are excluded. Raw reference recordings are also excluded. Source attribution, processing modifications and vector hashes are recorded in manifest.json and LICENSES/NOTICE.md.
Reproducibility and available files
This upload contains combined component weights, two selected speaker vectors, a manifest and license notices. It excludes inference/serving code, scheduling and orchestration, tokenizers, runtime configurations, datasets, reference recordings and training state.
The manifest's model/stack/selected_packages.json reference is an internal source-repository path, not a file available in this Hugging Face repository. The combined safetensors file alone is insufficient to reproduce the complete serving runtime. A publication date for the supporting code and configurations has not been announced.
Download
Download the Swarit Indic Talk model weights.
License and attribution
The repository metadata declares Apache-2.0. The combined artifact includes third-party components and speaker vectors that retain their applicable upstream terms, including MIT for the XLM-R encoder and CC BY 4.0 for the selected speaker-vector source material. See LICENSES/NOTICE.md and manifest.json for attribution and the component mapping.