Ayush Kumar Bar
Voice AI Engineer at Humyn Labs — speech, voice & LLM systems, from fine-tuning to real-time inference.

Building machines to understand humans.

• Working on BRIDGE ASR 2.0 ↗, an independent benchmark of 22 commercial ASR APIs on real conversational speech across 21 languages of the Global South — Indic, Latin American Spanish, Brazilian Portuguese and Vietnamese.
• Goes past the single score: a six-metric stack — WER, CER, semantic similarity, code-switch F1, phoneme-informed error rate and word information lost — after one shared normalisation for every provider.
• Evaluates spontaneous dual-speaker conversations (10–15 min, 44–48 kHz) with overlap and cross-talk, sliced across seven cohort dimensions such as region, noise level and speaker overlap.

• Fine-tuned and extended SparkTTS, MiraTTS, PiperTTS and QwenTTS with new language tokens and vocabularies for 22 Indic languages.
• Optimised SparkTTS inference with Triton Inference Server and vLLM to ~400 ms TTFT and 315 tokens/sec, profiling the Qwen2.5-0.5B → BiCodec encode/decode bottleneck.
• Fine-tuned Qwen3-ASR (0.6B, 1.7B) for 55 Indic languages across customer support, banking, finance and conversational speech.
• Served ASR through vLLM at 200 concurrent requests on a single A100, ~450 tokens/sec, RTFx 40.

• Fine-tuned Whisper, Indic Conformer and TTS models with LoRA (PEFT) and Seq2Seq Trainer on A100s, with Accelerate for distributed data parallel training.
• Led, built and deployed the real-time transcription service shunyalabs.ai/realtime ↗ on L40 GPUs — 250+ concurrent requests at low latency.
• Fine-tuned pingala-v1-en-verbatim ↗︎ to 2.94% WER, ahead of industry benchmarks for English verbatim transcription.
• Contributed to speech and voice AI research: dataset preparation, model optimisation and GPU utilisation.

• Built Stella 2.0, an agentic mental-health assistant: 15 autonomous agents across 60+ nodes on LangGraph, LangChain, Redis, Neo4j, Milvus and Kafka — live in the United We Care app ↗.
• Built a Clinical Copilot for healthcare professionals, using chain-of-thought prompting to reduce hallucinations and solve needle-in-a-haystack retrieval.
• Built real-time vector search APIs (L2, inner product, cosine) and moved service calls onto Kafka — field generation down 83%, from 30 to 5 minutes.