Skip to main content
Search

# Speech Recognition

14 related articles

Subscribe to updates
  1. arXiv

    Tencent Hunyuan releases Hy ASR 3.0: lowest aggregate error on public Mandarin, English and Cantonese benchmarks among evaluated systems

    Tencent Hunyuan released Hy ASR 3.0, a large-language-model-based speech recognition system covering Mandarin Chinese, English, Cantonese and 12 other Chinese dialect groups, with external context supplied as hotword lists, question-answer dialogues or summaries. On public Mandarin, English and Cantonese benchmarks it reaches aggregate error rates of 4.64%, 5.22% and 5.06%, each below the evaluated non-Hy systems; the Cantonese-only aggregate is about 11.0% lower than Qwen3 ASR Flash. The system also uses information inside the same utterance to disambiguate homophonic or phonetically similar keywords without a separate context input.

  2. Falcon-ASR takes on Arabic speech at 1.6B parameters: 20.92% average word error across six test sets, with Emirati dialect in focus

    Abu Dhabi's Technology Innovation Institute released Falcon-ASR, a 1.6-billion-parameter speech recognition model built for Arabic with particular attention to the Emirati dialect, and also covering English, French, Spanish and Portuguese. Across six Arabic test sets it reached an average word error rate of 20.92%, below the best published 23.17% in the leaderboard snapshot used for comparison; on an internal Emirati evaluation it recorded 22.73% word error and 10.19% character error, the lowest among the systems compared. The model also emits word-level timestamps that link each transcribed word to its position in the audio.

  3. Microsoft releases VibeVoice-ASR-Streaming: transcription emits speaker, time and content per chunk as speech arrives

    Microsoft released VibeVoice-ASR-Streaming, which turns the "who said what" transcription that previously required processing a whole audio segment into continuous output as speech arrives. The earlier VibeVoice-ASR handled 60-minute long-form audio in a single pass, producing structured transcriptions with speaker, timestamps and content within a 64K token length. The new model emits text once per chunk and supports customized hotwords and 10 languages.

  4. A hearing-aid chip folds filtering and recognition into one ASIC: 160,000 square micrometres and 1.4 milliwatts after synthesis

    Hearing aids must denoise and recognize speech within tight space and battery limits, and high-order finite impulse response filters cost area and power through many multiply-accumulate operations. This architecture places a memory-less, multiplier-less distributed-arithmetic FIR filter and a deep spectro neural network in one application-specific integrated circuit design, with inference using an approximate Wallace tree multiplier. Post-synthesis analysis reports 160,000 square micrometres, 1.4 milliwatts and about 2.1 milliseconds of hardware processing latency, while functional validation stays in MATLAB and prototype fabrication and real-time hardware validation remain future work.

  5. arXiv

    Speech LLMs stop pooling at a fixed frame rate: a learned boundary policy cuts English audio tokens by about a quarter

    ASSET trains a speech encoder's segment boundaries with reinforcement learning, reaching recognition close to a transcript-assisted character-boundary oracle on LibriSpeech with roughly 25% fewer audio tokens.

  6. arXiv

    A $5 chip runs open-vocabulary speech recognition: 13M-parameter model stays in flash, with 1.8–2.8 fewer errors than the vendor recognizer on the same chip

    Oído runs open-vocabulary English speech recognition entirely on a $5 ESP32-S3, a chip with no neural accelerator. Such microcontrollers previously managed only keyword spotting or a fixed command list, while a 13M-parameter Conformer-CTC occupies 14.0 MB at int8, too large for the 512 KB of on-chip SRAM and only partly placeable in PSRAM. The engine schedules execution around flash and PSRAM bandwidth, reaching 3.70%/8.23% word error rate on LibriSpeech test-clean/test-other, within 0.12 points of full precision and 1.8–2.8 fewer errors than the vendor recognizer on the same chip; measured on a board it runs at 1.97 times the audio duration on one core, not yet real time.

  7. arXiv

    Finnish speech models stop picking only clean broadcasts: 471K hours kept from 2.1M hours of raw audio by recognition confidence

    Finnish speech self-supervised pretraining no longer relies only on voice activity detection and language identification to pick data. Those pipelines ensured speech was present, not that it was clear; this pipeline uses ASR confidence as a proxy for pretraining data quality after voice activity detection and deduplication, keeping 471K hours from 2.1M hours of raw broadcast audio. On four benchmarks, confidence filtering gives 3-7% relative WER improvement over language-identification filtering despite a smaller retained training set.

  8. arXiv

    Speech decoding stops waiting token by token: a draft module reuses the target model's own features for 1.9x lossless speedup on Whisper

    Speech recognition decoding no longer has to wait token by token: a lightweight draft module that reuses the target model's acoustic and decoder features delivers up to 1.9x and 2.4x lossless speedups on Whisper and Conformer.

  9. TII releases 1.6B-parameter Falcon-ASR: 20.92% average WER across six Arabic test sets, 22.73% on internal Emirati evaluation

    The Technology Innovation Institute (TII) in Abu Dhabi released Falcon-ASR, a 1.6B-parameter speech recognition model focused on Arabic with particular attention to the Emirati dialect and also supporting English, French, Spanish and Portuguese; it reached 20.92% average word error rate across six Arabic test sets, better than the 23.17% best published result it cites, and 22.73% WER with 10.19% CER on an internal Emirati evaluation, the lowest among the systems compared, while also providing word-level timestamps.

  10. NVIDIA fine-tunes Nemotron 3.5 ASR on SADA and FLEURS, cutting Saudi Najdi and Hijazi WER from 55.05% to 29.96% while English also improves slightly

    This NVIDIA tutorial adapts a Cache-Aware FastConformer-RNNT multilingual streaming ASR model to Najdi and Hijazi Saudi dialects using minimal curation of SADA 2022 (retaining 103,559 of 125,490 utterances, 133.7 hours, 82.5%), a weighted replay mix of 90% Saudi speech with 7% English and 3% Arabic FLEURS, and duration bucketing, training for 12,000 steps in about 4.5 hours on two GPUs, which lowers Najdi+Hijazi test WER from 55.05% to 29.96% and CER from 31.63% to 12.18%, full SADA WER from 58.84% to 35.61%, while FLEURS English WER falls from 11.04% to 10.42% and Arabic WER from 12.67% to 11.41%; it also compares encoder unfreezing depths (top 6 at 33.42%, top 8 at 32.32%, all 24 layers at 29.96%) and inference-time settings ([56,13] attention context gains 1.