Towards Fully Open General Audio Intelligence
Speaker: Sreyan Ghosh - Google DeepMind

https://www.youtube.com/watch?v=n1QzhX_u-o0

watched this talk quickly, just couple points

1. Audio Flamingo injects audio features with cross-attention, not as tokens (not like speech LLMs - qwen, granite)
2. End to end was not successful (as usual, starts 48:10)

Tatiana left Apple and now works at Antropic

https://www.linkedin.com/in/tatiana-likhomanenko-36905925a

there are rumors about Antropic video/audio models too, hopefully we will see them soon.

https://huggingface.co/fishaudio/drama-3

35B params 3B active

Poor me reading all the Claude outputs literally everywhere. I hate it already. I barely understand the meaning. Guys, please do something about it. Help!

Few examples:

...The ElevenLabs rows are a control, not a complaint

...All 2,600 outputs remain in the evaluation. Twelve outputs reached the 512-token cap: 5 Kyrgyz, 3 Kazakh and 4 Russian (0.46% overall)

...so the choice of reward model is itself a robustness decision.

https://x.com/_josh_meyer_/status/2110043334182043658

How to kill a voice agent? Just turn on the TV

We benchmarked 14 voice agents with a TV playing in the background.

Task completion fell from 38.7% to 8.6%

Irodori is Japanese only but overall pretty advanced TTS. Here T5Gemma text encoder is particularly interesting

https://huggingface.co/Aratako/Irodori-TTS-v4-Large

Drifting is not a keyword anymore, we look for priming

https://x.com/Jaeyeon_Kim_0/status/2110744814274674851

https://github.com/JaeyeonKim01/priming

Dear colleagues,

We are happy to announce that The Sixth Workshop on NLP Applications to Field Linguistics (Field Matters 2027) will take place at EACL 2027 on March 9-14 in Athens, Greece (online participants are also welcome).

We accept papers on the following topics:
·         Application of NLP to field linguistics workflow;
·         The impact, benefits, and harms of NLP-assisted fieldwork;
·         Transfer learning for under-resourced language processing;
·         The use of fieldwork data to build NLP systems;
·         Modeling morphology and syntax of typologically diverse languages in the low-resource setting;
·         Speech processing for under-resourced languages;
·         Machine-readable field linguistic datasets and computational analysis of field linguistics datasets;
·         Using technology to preserve culture via language;
·         Improving ways of interaction with Indigenous communities.

Important dates
Direct Submission deadline: December 15, 2026
Pre-reviewed (ARR) submission deadline: December 22, 2026
Notification of acceptance: January 11, 2027
Camera-ready deadline: January 22, 2027
Workshop dates: March 9-14, 2027
You can find more information on the submission process and format requirements on our website.

For our 2027 edition, Field Matters will organise a Shared Task on Automatic Interlinear Glossing. More details can be found on the dedicated page.

If you have any questions, feel free to ask them!

COT-TTS: Audio Context-Aware Text-to-Speech with Chain-of-Thought Reasoning

https://luckybian.github.io/COT-TTS/

Similar ideas, TTS properties could be more predictable and explainable

https://huggingface.co/WhissleAI/steer-tts

Recently a good Arabic ASR was released:

https://huggingface.co/IFM/Jais-ASR

Good results on telephony, just a little behind the public leader (Openai GPT Audio 1.5) but still significantly behind custom finetuned models. For huge volume and smaller domain a simple conformer/zipformer can do much better than any advanced model but might fail on proper names and complex vocabulary.

Some other leaderboard results (not complete) here:

https://huggingface.co/spaces/elmresearchcenter/open_universal_arabic_asr_leaderboard

https://lnkd.in/p/eh-i56jw

https://opensvbench.xyz/

https://arxiv.org/abs/2610.13760

OpenSVBench: Diagnosing Speaker Verification Generalization Across Deployment Scenarios
Qituan Shangguan, Bosen Xu, Shuai Wang

Speaker verification (SV) systems now achieve very low error rates on standard public protocols, but those results do not fully describe deployment generalization. VoxCeleb-style evaluations remain useful anchors for English in-the-wild media speech. Scenario-specific evaluations and challenges are also valuable. However, neither setting directly shows whether the same system remains reliable across diverse deployment scenarios. We introduce OpenSVBench, a public benchmark for cross-scenario SV evaluation. Its fixed public core contains 26 trial sets organized into 11 scenario groups, with 9.15M verification trials, frozen target/non-target keys, and metadata for scenario-wise reporting. Around this core, the online leaderboard acts as a diagnostic evaluation interface rather than only a ranked list. It accepts score packages from new systems, evaluates them with frozen keys and common scoring, and reports trial-set breakdowns, scenario-group metrics, macro EER/minDCF, model-wise profiles, and scenario-wise comparisons. We evaluate 18 public checkpoints spanning x-vector, ECAPA-TDNN, CAMPPlus, ResNet and ERes variants, ReDimNet2, SAM-ResNet, and pretrained Transformer speaker recipes. The results show clear cross-scenario gaps: systems with similar macro EER can have different scenario profiles, strong performance in one scenario does not guarantee strong performance in another, and same-scenario trial sets can still expose source and protocol gaps.

Recently released MingTTS supposed to be good and just 0.5B params. And some other models too:

https://x.com/AntLingAGI/status/2112210811997540475

Today, we're open-sourcing 5 Ming-3.0 models across 3 speech families.

- Ming-3.0-tts: 0.5B voice design & paralinguistic cues; Tiny 7.9B A1.3B long-form speech.
- Ming-3.0-audio: Flash 126B A6.5B understanding & tools; Tiny 9.3B A2.7B efficient dialogue.
- Ming-3.0-asr: Tiny 8.6B A2B dialects & context.

Voice is somewhat robotic, not sure how it got good score

https://x.com/ArtificialAnlys/status/2112389524466929979/video/1

Not that the TTS is great but I find that using two personas for assistant is psychologically interesting thing

https://x.com/JonPTaylor/status/2112629213211828652

Lately there is a boom in tiny models, feels like everyone tired of growth. For example

https://github.com/yinruiqing/tiny-transducer

1M params transducer based on https://arxiv.org/abs/2101.06856

or this 1M TTS

https://huggingface.co/VantoraLabs/Vocetta2-1m

What is interesting the UTMOS reported is above 4.0 which assumes a good quality while in practice there is a clear metallic noise

https://huggingface.co/VantoraLabs/Vocetta2-1m/blob/main/samples/sample_3.wav

Feels like guys found a way to hack UTMOS getting good metrics without actual sound quality. No more UTMOS in our evaluation anymore then

https://huggingface.co/AuroraAI-Research/AuroraTTS-100K is another example

samples here

https://huggingface.co/AuroraAI-Research/AuroraTTS-100K/tree/main/samples

Encoder might be useful, not public though

https://arxiv.org/abs/2610.31292

This paper introduces Spark-Audio 1.0, a speech foundation model trained on domestically produced computing infrastructure. The model accepts text and speech inputs and generates text outputs. It supports multilingual (99 languages) and multi-dialect (202 dialects) speech recognition, speech translation, audio perception and understanding, and both text- and speech-based dialogue, marking a shift from merely ``hearing clearly'' to genuinely ``understanding'' speech. Architecturally, Spark-Audio 1.0 couples a 0.65B-parameter dense audio encoder with a 30B-A3B Mixture-of-Experts (MoE) large language model (LLM). The audio encoder is pretrained with representation-enhanced BEST-RQ and then refined through multi-task supervised training. The audio--LLM is trained through staged pretraining, followed by post-training that combines supervised fine-tuning with on-policy distillation. In total, the model is trained on 13 million hours of audio data together with large-scale text data. We evaluate Spark-Audio 1.0 on mainstream audio and text benchmarks, as well as on in-house test sets drawn from real-world application scenarios. Despite using substantially less audio training data than Qwen3.5-Omni-Flash, Spark-Audio 1.0 achieves comparable performance on most audio and text tasks, outperforms it in multilingual speech recognition, spoken dialogue, and real-world application scenarios, and approaches the performance of larger closed-source models. Nevertheless, there remains room for improvement in instruction following and complex audio reasoning. Finally, the successful training of Spark-Audio 1.0 on domestic hardware, including Huawei Ascend 910B and Hygon BW1000 chips, demonstrates the feasibility of a domestic technical pathway for building speech foundation models.

Some new Turkish models (and datasets/etc)

ASR

https://huggingface.co/erayyapagci/balbal-v1
https://huggingface.co/atasoglu/seda-v0.2

TTS

https://huggingface.co/kasparovabi/drama-3-tr-speaker

Time to go multimodal, great dataset for audio/video things

https://github.com/FreedomIntelligence/TalkVid