Canopy Labs

Serving Voice AI at $1/hr: Open-source, LoRAs, Latency, Load Balancing - Neil Dwyer, Gabber

Jul 31, 2025 · 16:09

Neil Dwyer, CTO of Gabber, details how his startup serves real-time voice AI for under $1 per hour using open-source Orpheus TTS, LoRAs for emotive voice cloning, and vLLM with FP8 dynamic quantization to achieve 95-105 tokens per second on L40S GPUs. He explains the critical latency challenge of 'head of line silence' in Orpheus (600ms in default voices) and how fine-tuning LoRAs reduces it to ~100ms, fitting within a 1.5-second budget for real-time conversation. Gabber batches multiple LoRA generations per GPU and uses a consistent hash ring for load balancing across servers, enabling popular clones to be dynamically replicated. The talk argues that with current open-source tools, building affordable consumer voice AI is accessible to small teams.

Optimizing inference for voice models in production - Philip Kiely, Baseten

Jul 1, 2025 · 15:13

Philip Kiely of Baseten shows how open-source TTS models like Orpheus TTS, built on a LLaMA 3.2 3B backbone, can be optimized for production inference using LLM tooling such as TensorRT-LLM and FP8 quantization, achieving time-to-first-byte (TTFB) under 150 milliseconds and supporting 16–24 simultaneous streams on a single half-H100 GPU. He argues that because TTS models are architecturally similar to LLMs, techniques like dynamic batching, KV-cache quantization, and Torch Compile on the audio decoder can dramatically reduce latency and increase concurrency. He emphasizes that for real-time voice agents, the goal is not raw tokens per second (real-time requires only 83 TPS for Orpheus) but low TTFB and high throughput to minimize GPU spend. However, Kiely warns that non-runtime factors—such as client code using sequential requests without session reuse or sending traffic to distant data centers—can easily add back the milliseconds saved at the model level, and that infrastructure connecting listening, thinking, and talking pipelines is often the dominant source of latency.

Text-to-Speech Data Preparation and Fine-tuning Workshop - Ronan McGovern

Jun 3, 2025 · 34:00

Ronan McGovern walks through fine-tuning Sesame's CSM-1B text-to-speech model on a specific voice, using a YouTube video as the data source. He explains token-based TTS models, including how audio is represented via codebooks and how CSM-1B uses a main transformer for zeroth tokens and a secondary transformer for 31 hierarchical tokens. The workshop covers data preparation: downloading audio with yt-dlp, transcribing with Whisper Turbo, manually correcting the transcript, and splitting audio into 30-second chunks (41 clips from a 30-minute video). Fine-tuning uses Unsloth with LoRA adapters (rank 32, alpha 16) on linear layers, training for one epoch with a batch size of 2 and virtual size of 8, reducing loss from ~6.34 to ~3.72. Evaluation compares zero-shot inference (random speaker), voice cloning (closer but imperfect), and fine-tuned plus cloning (best result, producing an Irish-accented voice with natural errors). McGovern recommends 50+ 30-second clips for noticeable effect and notes that combining fine-tuning with voice cloning yields good performance even with limited data.