{"slug":"speech-recognition","label":"Speech Recognition","item_count":3,"day_count":3,"source_count":3,"first_seen":"2026-10-28T19:30:01+00:00","last_updated":"2026-11-10T17:59:28+00:00","generated_at":"2026-11-12T00:15:17.843192+00:00","sources":["arxiv_cs_cl","arxiv_cs_lg","arxiv_llm_reliability"],"days":[{"date":"2026-10-28","items":[{"title":"Streaming Speech Recognition on a Microcontroller: a 65.7M-Parameter Transducer in 2 and 4 Bits on the ESP32-P4","url":"http://arxiv.org/abs/2610.38739v1","source":"arxiv_cs_lg","type":"paper","summary_1line":"We run Kroko, a 65.7M-parameter streaming Zipformer2 transducer for English, on the ESP32-P4, a microcontroller with two 400 MHz RISC-V cores, a 128-bit SIMD extension (PIE) and 32 MB of PSRAM, but no neural accelerator. Without any training, the encoder's 290 matrices are stored","sid":"64d5612d6441ca2f","published":"2026-10-28T19:30:01+00:00","editor_note":"The deployment-footprint data point: low-bit storage of the encoder's 290 matrices without any training, on a microcontroller."}]},{"date":"2026-10-30","items":[{"title":"From Transcription to Reformulation: Defining and Benchmarking Grounded Generative Speech Recognition","url":"http://arxiv.org/abs/2610.41292v1","source":"arxiv_llm_reliability","type":"paper","summary_1line":"Speech language models have expanded automatic speech recognition (ASR) from transcription to general speech-to-text generation. Yet existing benchmarks still assume a single ground-truth transcript per utterance, an...","why_it_matters":"Matches feed focus: evaluation.","sid":"e829938c6d7f203a","published":"2026-10-30T05:53:56+00:00","editor_note":"The evaluation angle: questions the single-ground-truth-transcript assumption behind current ASR benchmarks."}]},{"date":"2026-11-10","items":[{"title":"Earlybird: Predictive Streaming Speech Recognition in 29M Parameters","url":"http://arxiv.org/abs/2611.14421v1","source":"arxiv_cs_cl","type":"paper","summary_1line":"Streaming speech recognizers output a word only after the speaker has finished it, typically 90-400 ms later. For a voice agent every millisecond counts: the sooner the words are outputted, the sooner the agent can st...","why_it_matters":"Matches feed focus: agent.","sid":"1cd37d39d82ad148","published":"2026-11-10T17:59:28+00:00","editor_note":"The latency angle: frames word-output delay as the cost a voice agent pays before it can start responding."}]}],"editorial":{"tldr":"This is a topical cluster of three independent arXiv papers on speech recognition, not one unfolding event. A late-October paper ran Kroko, a 65.7M-parameter streaming Zipformer2 transducer, at 2 and 4 bits on an ESP32-P4 microcontroller with no neural accelerator, and a second argued that ASR benchmarks still assume a single ground-truth transcript per utterance.","stale":false,"whats_new":"Earlybird, a 29M-parameter predictive streaming recognizer, targets the 90-400 ms that streaming ASR typically waits after a speaker finishes a word before outputting it, a delay it frames as a voice-agent cost.","why_it_matters":"The three papers touch separate parts of a voice-agent stack: where the recognizer can run, how soon it emits words, and whether a single-reference benchmark can score speech models that generate rather than transcribe.","take_for_builders":"Treat all three as single-paper claims: before changing a voice pipeline, measure your own recognizer's word-finalization delay, since that is the number Earlybird's 90-400 ms framing says to compare against.","beats":[{"kicker":"FOOTPRINT","tone":"launch","headline":"A 65.7M-parameter streaming transducer runs on an ESP32-P4 in 2 and 4 bits","summary":"The target chip has two 400 MHz RISC-V cores, a 128-bit SIMD extension and 32 MB of PSRAM, with no neural accelerator.","sids":["64d5612d6441ca2f"]},{"kicker":"EVALUATION","tone":"neutral","headline":"A benchmark proposal for grounded generative speech recognition","summary":"The paper argues that speech language models moved ASR from transcription to general speech-to-text generation while benchmarks kept one reference transcript per utterance.","sids":["e829938c6d7f203a"]},{"kicker":"LATENCY","tone":"now","headline":"Earlybird predicts words early to cut streaming lag in 29M parameters","summary":"It starts from the observation that streaming recognizers output a word 90-400 ms after the speaker finishes it.","sids":["1cd37d39d82ad148"]}],"open_questions":["What accuracy does the 2-bit and 4-bit Kroko encoder lose against the full-precision model, and does it run in real time on the ESP32-P4?","How much of the 90-400 ms output delay does Earlybird remove, and at what word-error cost when a prediction is wrong?","Do existing speech language models rank differently under a grounded generative benchmark than under single-reference WER?","Does any of the three papers ship code or weights that others can reproduce?"],"generated_at":"2026-11-12T00:13:57Z"}}