AnySpeechRecognizer, ElevenLabs Cue, DAYJOB: Tax

Three items today, and they share a theme: the unglamorous parts of AI work, such as transcribing, timing and doing a professional job end to end, are where the measurable gaps are.

AnySpeechRecognizer and AnyDecisionModel

Hugging Face published two Swift packages for Apple platforms. AnySpeechRecognizer mirrors Apple's SpeechAnalyzer and SpeechTranscriber API but lets you pick the engine, so switching can be as small as changing an import. The engines include NVIDIA's Parakeet and Alibaba's Qwen3-ASR running locally on Core ML, and the post adds a speaker diarizer built on NVIDIA's Streaming Sortformer that tracks up to four speakers.

The measurements come from an M4 Max Mac Studio on LibriSpeech files of 3 to 8 minutes. Parakeet Unified reaches 1.43% word error rate on test-clean from a file, against 2.12% for Apple's own transcriber. On AMI meeting recordings, where people talk over each other, every engine they tested stayed above 23%. The authors say these numbers come from one machine and a few datasets, and that they tested on macOS 27 only, not an iPhone.

AnyDecisionModel is the second package. Instead of generating text and parsing it, it reads a probability for each allowed answer from the model, so a question like "Which team should handle this ticket?" returns a typed choice with probabilities and a confidence value. It has six backends, including MLX, llama.cpp and Ollama, and supports Cloudflare's Clef decision-head models. The Core ML conversions list their licenses on the Hub, from CC BY 4.0 to Apache 2.0, so check them before shipping.

ElevenLabs Cue

ElevenLabs described Cue, the forced alignment model that places every character of a transcript on an audio timeline. It sits behind the Scribe speech-to-text family and is also available through a Forced Alignment API.

The company reports an average timing error of 179 ms on full recordings, while the open-source tools it tested ranged from 283 ms to more than 6,000 ms. On English recordings with background noise the figure is 57 ms against 65 to 920 ms. Cue handles recordings of any length in one pass, so there is no chunking error, and it aligns an hour of audio in about half a second on ElevenLabs' production GPU. The next fastest tool, Qwen3-ForcedAligner, takes nearly 2 seconds.

The benchmark is ElevenLabs' own, built from real requests and covering 29 languages, and it scores character-level output with word-level labels, which the team says likely understates Cue. On songs from the JamendoLyrics dataset it places 92.7% of words within 0.3 seconds of the label.

DAYJOB: Tax

Surge AI released a benchmark of 50 tax environments across eight practice areas, with 1,789 files including filed returns, K-1s, ledgers and a photographed handwritten receipt. Agents get a client request and must produce returns, workpapers or advice, including issues the client never raised.

Claude Sonnet 5.5 leads at a 10.8% pass rate, followed by Claude Fable 5.1 at 9.6% and Claude Opus 5.5 at 9.2%, measured as pass@1 at a 95% rubric threshold. The case study that stuck with me: an agent correctly compared two closing dates for a condo sale but missed that waiting past August 2027 could cost a tax exclusion worth up to $250,000.

What I'd watch next

I want to see whether Cue-style alignment shows up in open models, and whether the Parakeet numbers hold on an iPhone. For DAYJOB, the interesting move is the next release of the leaderboard, to see if pass rates climb past low double digits.