Neosophie

Blog

Articles about Neosophie's product development and technology.

#asr #books #diarization #finetune #llm #metrics #oss #rag #tts

April 8, 2026asr finetune

How I Finetuned IBM Granite Speech 1B on Japanese Audio and Improved CER from 0.37 to 0.14

I finetuned IBM Granite Speech (`granite-4.0-1b-speech`) on 100 hours of Japanese speech data and reduced CER from 0.37 to 0.14. The official script's Projector+LoRA-only training has a ceiling on accuracy gains. The key breakthrough was additionally training `lm_head` and the last 8 layers of the Language Model. The result matches Qwen3-ASR-1.7B (CER 0.14) with only 1B parameters.

Neosophie Team

March 23, 2026

Japanese ASR Model Comparison 2026: Whisper, Qwen3, Voxtral & ReazonSpeech Benchmarked on RTX5090

Benchmarking 8 Japanese ASR models under identical conditions on an RTX5090, qwen/qwen3-asr-1.7b (WER: 0.1899) and whisper (WER: 0.2099) clearly lead in both accuracy and stability. For speed-critical workloads, parakeet-tdt-0.6b-v3 (RTF: 0.002) is the fastest by a wide margin, while reazonspeech-espnet-v2 is the strongest choice for Japanese broadcast/media domains. To ensure fair comparison across models, WER was computed using MeCab morphological tokenization combined with punctuation stripping before scoring.

Neosophie Team

March 20, 2026tts

Open-Source TTS Model Showdown 2026: Hands-On Testing of Qwen3-TTS, Chatterbox, Fish Speech, CosyVoice & IndexTTS2

A hands-on comparison of five open-source TTS models, with a focus on commercial license pitfalls in Fish Speech and IndexTTS2 — covering features, licenses, and working code all in one place.

Neosophie Team

March 19, 2026tts

I Tested Zero-Shot Voice Cloning with Emotion Control in OpenVoice — 8 Styles from a 14-Second Reference

OpenVoice V1 is a zero-shot voice cloning library that extracts a speaker's tone color from as little as 14 seconds of reference audio, then synthesizes speech in 8 emotional styles: `whispering`, `shouting`, `excited`, `cheerful`, `terrified`, `angry`, `sad`, and `friendly`. This post covers the code I actually ran, what I heard, and the gotchas I hit along the way.

Neosophie Team

March 12, 2026books

5 Books a Working AI Engineer Actually Read for Real-World Accuracy Improvement and Product Development

The reason you can call AI APIs but still can't improve accuracy is a lack of understanding of what's happening inside the model. These 5 books — all O'Reilly titles written by practitioners at the forefront of the field, including Hugging Face engineers and Chip Huyen — are the standard reading list for ML engineers worldwide.

Neosophie Team

March 12, 2026rag

Replaced Embedding RAG with PageIndex on My Own Site — Here's What Actually Happened

After hitting a wall with hybrid BM25 + Embedding + Rerank search (wrong retrievals, high infra cost, Japanese embedding model limitations), I implemented **PageIndex** — a vector-free RAG architecture — as the AI assistant for my company site. The result: no VectorDB, no embedding model, no Reranker needed. Two-stage LLM-based filtering via a URL tree navigated to the correct pages for queries like *"What products do you have?"* and *"What ASR-related blog posts exist?"* — zero vector math involved.

Neosophie Team

March 12, 2026diarization metrics

Speaker Diarization Evaluation Metrics: Beyond DER — A Practical Diagnostic Guide

Evaluating speaker diarization systems requires choosing the right metric for what you actually want to measure. DER is the industry standard, but it cannot diagnose *why* a system fails. To systematically improve a real system, **combine DER with Confusion Matrix, Purity/Coverage, and Boundary Error** for proper root-cause analysis. This article covers calculation formulas, working Python code, and real experimental data — not just textbook definitions.

Neosophie Team

March 12, 2026llm finetune

Suppressing LLM Repetition Hallucinations with a Custom Logits Processor in Qwen

Fine-tuning Qwen for structured JSON output can trigger repetition hallucinations where the model loops the same phrase indefinitely. Standard parameters like no_repeat_ngram_size suppress this globally, causing unintended side effects across the entire output. This article implements a custom Transformers LogitsProcessor that applies repetition control exclusively inside the target JSON field, eliminating loops without breaking the surrounding structure.

Neosophie Team

March 12, 2026diarization

Speaker Diarization Benchmark: NeMo vs VibeVoice vs Pyannote — Which Model Actually Works?

I tested three major speaker diarization frameworks — **NVIDIA NeMo** (Clustering, MSDD, Sortformer), **Microsoft VibeVoice**, and **pyannote.audio** — against real-world audio: a 10-minute drama clip with 14 speakers and two 10-minute debate clips with 6–10 speakers.

Neosophie Team

March 12, 2026oss

Free Offline Transcription App for Mac: How I Built Local Whisper with whisper.cpp + CoreML

I built a fully offline, free transcription app for Mac called **Local Whisper** that captures both microphone input and system audio (including your meeting partner's voice). It runs whisper.cpp with CoreML acceleration on Apple Silicon, achieving near-real-time transcription without sending any data to the cloud. This post covers the architecture, the real-time optimization tricks, and the system audio capture method that took the most effort to crack.

Neosophie Team