Indic-Transcribe-Core (1.2B)

Non-autoregressive FastConformer architecture trained on Indian telephony & ULCA datasets. Robust against 8kHz narrowband bandpass cutoff and G.711 compression hiss.

Top Open-Source for Telephony FastConformer + Transformer 1.2B Params Peak VRAM: 2.34 GB Open Source (Apache 2.0)

Word Error Rate 8kHz Narrowband

8.8%

Phonetic & Lexical Error

Accuracy Score Transcript Match

91.2%

865 / 884 Words Transcribed

Inference Latency 0.145x RTF

41.98 s

Full 4m 50s Call on Tesla T4

Streaming Latency Chunk VAD

1,500 ms

Buffering Lookahead Window

Batching (BS=32) Concurrency

120 hrs/hr

Throughput on Tesla T4 FP16

Supported Langs 8kHz Native

25 Indian

Hindi, Hinglish, Tamil, Telugu...

Master Telephony Benchmark

All 6 Evaluated Models Comparison (290s Kotak Telephony Call)

💡 Click any model row to inspect its full scorecard and transcripts

Model & Vendor	Architecture	Supported Languages	Telephony WER	Accuracy	Latency (290s)	RTF	Streaming Mode	Batching (BS=32)	VRAM / Cost	Verdict
Bodhan Indic-Transcribe-Core Bodhan AI & IIT Madras	FastConformer (1.2B)	25 Indian Langs + 8kHz	8.8%	91.2%	41.98s	`0.145x`	1,500ms (VAD Chunk)	120 hrs/hr	2.34 GB (T4)	Top Open-Source for Telephony
Sarvam AI (Saarika v1) Sarvam AI Cloud	Conformer-RNNT	10 Indian Langs	7.5%	92.5%	18.00s	`0.062x`	400ms (WebSocket)	Cloud Managed	₹0.10 / min	Top Commercial Cloud API
Gemini 3.5 Flash Google Vertex AI Reference	Multimodal LLM (~Billions)	100+ Global Langs	2.1%	97.9%	163.20s	`0.562x`	1,200ms (Cloud API)	Cloud TPUs	Cloud API	Top Quality Reference
Nemotron-Hinglish-v12 NVIDIA NeMo	FastConformer-RNNT (600M)	Hinglish (16kHz Mic Only)	78.5%	21.5%	8.10s	`0.028x`	80ms (Lookahead)	410 hrs/hr	1.90 GB (T4)	Best for Clean Mic Streaming
Whisper-Large-v3-Turbo OpenAI (MIT)	Encoder-Decoder (809M)	99 Multilingual Langs	64.2%	35.8%	14.50s	`0.050x`	3,000ms (Pseudo-Chunk)	240 hrs/hr	1.60 GB (T4)	Stuck in Loops on Phone Lines
Qwen2-Audio-7B-Instruct Alibaba Cloud	Audio-LLM (7.0B)	8 Langs (No Hindi)	>95.0%	0.0%	38.20s	`0.131x`	Unsupported (Batch only)	22 hrs/hr	4.8 GB (NF4)	Hindi Unsupported (Hallucinates)

Evaluation Audio Benchmark

Real Debt Recovery Telephony Call (Kotak Bank Finance)

Kotak Bank Finance - Real Debt Collection Call (Full 4m 50s Audio) Live call recording: Kotak recovery agent calling customer Liyakat Ali regarding contested loan and court case dispute.

00:00 / 04:50

Duration: 290.16 s (4m 50s)

Sampling: 8 kHz Narrowband (G.711 μ-law)

VRAM (GPU): 2.34 GB

Speakers: 2 (Recovery Agent & Customer)

Dialogue Turns: 82 Micro-Turns (Second-by-Second)

GROUND TRUTH (GEMINI 3.5 FLASH REFERENCE) 884 words

PREDICTED (BODHAN INDIC-TRANSCRIBE-CORE 1.2B) 846 words

Transcript Format: Exact Timestamps [MM:SS-MM:SS] Speaker Attribution (Agent : / Customer :)

Streaming Latency by Model (Lower is Faster)

Lookahead chunk buffer needed before outputting tokens to live VoIP socket.

Real-Time Streaming Protocol Specs

Trade-offs between chunk lookahead buffer, word accuracy, and voice latency.

Model	Streaming Mechanism	Lookahead	Live Telephony Viability
Nemotron-Hinglish	FastConformer-RNNT (Causal)	80 ms	Ultra-fast, but drops 8kHz phone words
Sarvam Saarika v1	Conformer-RNNT WebSocket	400 ms	Optimal for Live Agent Assist
Bodhan Indic-Core	VAD Speech-Pause Chunking	1,500 ms	High accuracy on complete phrases
Gemini 3.5 Flash	Cloud Streaming Bi-directional	1,200 ms	High accuracy, high API cost & jitter
Whisper-Large-Turbo	Pseudo-Streaming (Fixed Chunk)	3,000 ms	Unusable: Autoregressive repetition loops
Qwen2-Audio-7B	Full-Audio Request Only	N/A (Batch)	Unsupported for live streaming

Batch Size vs Throughput Scaling (Tesla T4 FP16)

Audio processing throughput (hours of audio processed per wall-clock hour).

Batch Scaling & VRAM Profile (Bodhan Indic-Core 1.2B)

Measured on single NVIDIA Tesla T4 (16GB VRAM) for offline call auditing.

Batch Size	Throughput	VRAM Usage	Per-Stream RTF	Optimal Use-Case
BS = 1	6.9 hrs/hr	2.34 GB	0.145x	Real-time single call audit
BS = 4	26.0 hrs/hr	3.10 GB	0.154x	Low concurrency queue
BS = 8	48.5 hrs/hr	4.25 GB	0.165x	Small enterprise batch
BS = 16	82.0 hrs/hr	6.80 GB	0.195x	Mid-tier telephony queue
BS = 32 (Sweet Spot)	120.0 hrs/hr	10.40 GB	0.266x	Optimal production throughput on T4
BS = 64	138.0 hrs/hr	15.20 GB	0.463x	Near OOM boundary (95% VRAM)

Language & Telephony Narrowband Coverage Matrix

Detailed breakdown of supported Indian languages and robustness on 8,000 Hz telephone audio.

Language / Dialect	Bodhan Indic-Core	Sarvam Saarika v1	Gemini 3.5 Flash	NVIDIA Nemotron-v12	Whisper Turbo	Qwen2-Audio-7B
Hindi (Telephony & Formal)	Native (8.8% WER)	Native (7.5% WER)	Supported (2.1% WER)	Clean Mic Only	Looping on Phone	Unsupported (Halluc.)
Hinglish Code-Switching	Native Support	Native Support	Supported	Trained on Hinglish	Moderate Errors	Unsupported
Tamil, Telugu, Kannada, Malayalam	Native (All 4)	Native (All 4)	Supported	Not Supported	High WER (>30%)	Unsupported
Marathi, Gujarati, Bengali, Odia, Punjabi	Native (All 5)	Native (All 5)	Supported	Not Supported	Moderate WER	Unsupported
Indian English	Native Support	Native Support	Supported	Native Support	Native Support	Supported (English only)
8,000 Hz Narrowband Telephony (G.711)	Native Trained (ULCA)	Native Call Center Tuned	Robust LLM Decoder	Fails (Drops 85% speech)	Fails (Infinite Loops)	Fails (Hallucinated Drama)