https://loudness.ai/research/Explore speech and audio experiments with clear methods, measured results and reproducible study data.enhttps://loudness.ai/research/piper-tts-voices-ranked/https://loudness.ai/research/piper-tts-voices-ranked/Thu, 15 Oct 2026 00:00:00 GMTWe tested every Piper text-to-speech voice that a business may use: 75 official voice files in 27 languages plus three community models trained from scratch, 1,683 voices in all once each speaker of a multi-speaker model is counted. Each voice read the same three sentences per language and received a naturalness score from UTMOS22 and a clarity score from Whisper large-v3, combined into an overall score and ranked within its language. The best voices are English; a community model trained from scratch, Clean 100, matched the best official voice (96.4 against 96.2) while running four to six times faster on a CPU. Speaker choice inside a model changed scores by up to 46.3 points. 102 of the 177 official files were excluded because their licence, or the Lessac voice they were fine-tuned from, forbids commercial use. An interactive explorer lets developers filter the voices by language, gender, licence and score, listen to samples and download models from their original source.https://loudness.ai/research/piper-tts-text-normalization/https://loudness.ai/research/piper-tts-text-normalization/Thu, 15 Oct 2026 00:00:00 GMTPiper is a fast neural text-to-speech engine that runs on ordinary processors, but it reads text exactly as written. Prices, dates, phone numbers, abbreviations and units are often read literally: “$5.50” becomes “dollar five point five zero”, and “Dr. Smith” ends a sentence after “Doctor”. We built a text preparation layer for Piper that rewrites these forms while remembering, for every spoken word, the written word it came from. It also sets pauses between sentences, lines and paragraphs, repairs British pronunciations in some American English voices, and gives exact word timings for subtitles from Piper’s own phoneme durations. On 84 English and Spanish test sentences we wrote, the phoneme error rate fell from 32.6% to 0.9%, and none of 19 control sentences changed. On the held-out half of NVIDIA NeMo’s public test cases, which we did not write, the phoneme error rate in English was 41.0% for unprocessed text, 28.5% for a typical number-to-words script and 15.8% with our layer; in Spanish it was 64.7%, 56.2% and 43.2%. The code, test sentences and results are available under the Apache License 2.0.https://loudness.ai/research/parakeet-whisper-cpu-benchmark/https://loudness.ai/research/parakeet-whisper-cpu-benchmark/Wed, 14 Oct 2026 00:00:00 GMTWe evaluated three Parakeet TDT 0.6B v3 configurations and two quantized Whisper variants on identical English reading audio using four CPU threads. On a 310.425-second recording, Parakeet FP32 with shorter chunks took 53.58 seconds with 2.81% word error rate (23 edits across 818 reference words) and 2.87 GiB peak process memory. Whisper Turbo INT8 took 343.26 seconds with 2.20% WER; Distil-Whisper INT8 took 279.10 seconds with 2.57% WER. Default Parakeet INT8 chunks omitted substantial speech; shorter chunks reduced WER from 39.12% to 4.65%. These single-request observations favor FP32 Parakeet for this CPU and sample, while the tested Whisper variants made fewer word edits. We publish all 18 completed runs, transcripts, fixture audio, references, checksums and reproduction settings.