Srikanth Ronanki

Amazon AGI, Aaron Langford, Aayush Shah et al. · amazon-science

We present Amazon Nova, a new generation of state-of-the-art foundation models that deliver frontier intelligence and industry-leading price performance. Amazon Nova Pro is a highly-capable multimodal model with the best combination of accuracy, speed, and cost for a wide range of tasks. Amazon Nova Lite is a low-cost multimodal model that is lightning fast for processing images, video, documents and text. Amazon Nova Micro is a text-only model that delivers our lowest-latency responses at very low cost. Amazon Nova Canvas is an image generation model that creates professional grade images with rich customization controls. Amazon Nova Reel is a video generation model offering high-quality outputs, customization, and motion control. Our models were built responsibly and with a commitment to customer trust, security, and reliability. We report benchmarking results for core capabilities, agentic performance, long context, functional adaptation, runtime performance, and human evaluation.

Saket Dingliwal, Monica Sunkara, Sravan Bodapati et al.

End-to-end speech recognition models trained using joint Connectionist Temporal Classification (CTC)-Attention loss have gained popularity recently. In these models, a non-autoregressive CTC decoder is often used at inference time due to its speed and simplicity. However, such models are hard to personalize because of their conditional independence assumption that prevents output tokens from previous time steps to influence future predictions. To tackle this, we propose a novel two-way approach that first biases the encoder with attention over a predefined list of rare long-tail and out-of-vocabulary (OOV) words and then uses dynamic boosting and phone alignment network during decoding to further bias the subword predictions. We evaluate our approach on open-source VoxPopuli and in-house medical datasets to showcase a 60% improvement in F1 score on domain-specific rare words over a strong CTC baseline.

Sai Muralidhar Jayanthi, Devang Kulshreshtha, Saket Dingliwal et al. · mila

Personalization of automatic speech recognition (ASR) models is a widely studied topic because of its many practical applications. Most recently, attention-based contextual biasing techniques are used to improve the recognition of rare words and domain specific entities. However, due to performance constraints, the biasing is often limited to a few thousand entities, restricting real-world usability. To address this, we first propose a "Retrieve and Copy" mechanism to improve latency while retaining the accuracy even when scaled to a large catalog. We also propose a training strategy to overcome the degradation in recall at such scale due to an increased number of confusing entities. Overall, our approach achieves up to 6% more Word Error Rate reduction (WERR) and 3.6% absolute improvement in F1 when compared to a strong baseline. Our method also allows for large catalog sizes of up to 20K without significantly affecting WER and F1-scores, while achieving at least 20% inference speedup per acoustic frame.

Veera Raghavendra Elluru, Devang Kulshreshtha, Rohit Paturi et al. · mila

Spoken language understanding systems using audio-only data are gaining popularity, yet their ability to handle unseen intents remains limited. In this study, we propose a generalized zero-shot audio-to-intent classification framework with only a few sample text sentences per intent. To achieve this, we first train a supervised audio-to-intent classifier by making use of a self-supervised pre-trained model. We then leverage a neural audio synthesizer to create audio embeddings for sample text utterances and perform generalized zero-shot classification on unseen intents using cosine similarity. We also propose a multimodal training strategy that incorporates lexical information into the audio representation to improve zero-shot performance. Our multimodal training approach improves the accuracy of zero-shot intent classification on unseen intents of SLURP by 2.75% and 18.2% for the SLURP and internal goal-oriented dialog datasets, respectively, compared to audio-only training.

Goeric Huybrechts, Srikanth Ronanki, Xilai Li et al.

Conformer-based end-to-end models have become ubiquitous these days and are commonly used in both streaming and non-streaming automatic speech recognition (ASR). Techniques like dual-mode and dynamic chunk training helped unify streaming and non-streaming systems. However, there remains a performance gap between streaming with a full and limited past context. To address this issue, we propose the integration of a novel dynamic contextual carry-over mechanism in a state-of-the-art (SOTA) unified ASR system. Our proposed dynamic context Conformer (DCTX-Conformer) utilizes a non-overlapping contextual carry-over mechanism that takes into account both the left context of a chunk and one or more preceding context embeddings. We outperform the SOTA by a relative 25.0% word error rate, with a negligible latency impact due to the additional context embeddings.

Dhanush Bekal, Sundararajan Srinivasan, Sravan Bodapati et al.

In this work, we define barge-in verification as a supervised learning task where audio-only information is used to classify user spoken dialogue into true and false barge-ins. Following the success of pre-trained models, we use low-level speech representations from a self-supervised representation learning model for our downstream classification task. Further, we propose a novel technique to infuse lexical information directly into speech representations to improve the domain-specific language information implicitly learned during pre-training. Experiments conducted on spoken dialog data show that our proposed model trained to validate barge-in entirely from speech representations is faster by 38% relative and achieves 4.5% relative F1 score improvement over a baseline LSTM model that uses both audio and Automatic Speech Recognition (ASR) 1-best hypotheses. On top of this, our best proposed model with lexically infused representations along with contextual features provides a further relative improvement of 5.7% in the F1 score but only 22% faster than the baseline.

Yufeng Du, Phillip Harris, Minyang Tian et al.

We identify intrinsic limitations of Rotary Positional Embeddings (RoPE) in Transformer-based long-context language models. Our theoretical analysis abstracts away from the specific content of the context and depends only on its length. We prove that as context length increases, RoPE-based attention becomes unpredictable and loses two properties that are central to its effectiveness. First, it loses its locality bias: RoPE is no more likely to favor nearer positions than substantially farther ones. Second, it loses consistency in token relevance: a key vector that receives a higher attention score than an alternative at one position may receive a lower score at another. In both cases, the probability of failure approaches 0.5, no better than random guessing. We further prove that the attention score can remain unchanged when a key token is moved to a different position, or even replaced by a different token, indicating a failure to distinguish positions or tokens. Adjusting the RoPE base trades off distinguishing positions against distinguishing tokens but cannot preserve both at the same time. Increasing the RoPE base hyperparameter, a common practice in today's long-context models, helps distinguish different tokens, but inevitably sacrifices the ability to distinguish positions. Our empirical analysis shows that multi-head, multi-layer architectures are insufficient to overcome these limitations. Our findings suggest that fundamentally new mechanisms for encoding position and token order may be needed in future Transformer long-context language models.

Zeyu Liu, Souvik Kundu, Lianghao Jiang et al.

Although transformer architectures have achieved state-of-the-art performance across diverse domains, their quadratic computational complexity with respect to sequence length remains a significant bottleneck, particularly for latency-sensitive long-context applications. While recent linear-complexity alternatives are increasingly powerful, effectively training them from scratch is still resource-intensive. To overcome these limitations, we propose LAWCAT (Linear Attention with Convolution Across Time), a novel linearization framework designed to efficiently transfer the capabilities of pre-trained transformers into a performant linear attention architecture. LAWCAT integrates causal Conv1D layers to enhance local dependency modeling and employs normalized gated linear attention to improve generalization across varying context lengths. Our comprehensive evaluations demonstrate that, distilling Mistral-7B with only 1K-length sequences yields over 90\% passkey retrieval accuracy up to 22K tokens, significantly extending its effective context window. Similarly, Llama3.2-1B LAWCAT variant achieves competitive performance on S-NIAH 1\&2\&3 tasks (1K-8K context length) and BABILong benchmark (QA2\&QA3, 0K-16K context length), requiring less than 0.1\% pre-training tokens compared with pre-training models. Furthermore, LAWCAT exhibits faster prefill speeds than FlashAttention-2 for sequences exceeding 8K tokens. LAWCAT thus provides an efficient pathway to high-performance, long-context linear models suitable for edge deployment, reducing reliance on extensive long-sequence training data and computational resources. Code is released at: https://github.com/zeyuliu1037/LAWCAT

Nilaksh Das, Saket Dingliwal, Srikanth Ronanki et al. · amazon-science

Large language models (LLMs) have shown incredible proficiency in performing tasks that require semantic understanding of natural language instructions. Recently, many works have further expanded this capability to perceive multimodal audio and text inputs, but their capabilities are often limited to specific fine-tuned tasks such as automatic speech recognition and translation. We therefore develop SpeechVerse, a robust multi-task training and curriculum learning framework that combines pre-trained speech and text foundation models via a small set of learnable parameters, while keeping the pre-trained models frozen during training. The models are instruction finetuned using continuous latent representations extracted from the speech foundation model to achieve optimal zero-shot performance on a diverse range of speech processing tasks using natural language instructions. We perform extensive benchmarking that includes comparing our model performance against traditional baselines across several datasets and tasks. Furthermore, we evaluate the model's capability for generalized instruction following by testing on out-of-domain datasets, novel prompts, and unseen tasks. Our empirical experiments reveal that our multi-task SpeechVerse model is even superior to conventional task-specific baselines on 9 out of the 11 tasks.

Do June Min, Karel Mundnich, Andy Lapastora et al.

One common approach for question answering over speech data is to first transcribe speech using automatic speech recognition (ASR) and then employ text-based retrieval-augmented generation (RAG) on the transcriptions. While this cascaded pipeline has proven effective in many practical settings, ASR errors can propagate to the retrieval and generation steps. To overcome this limitation, we introduce SpeechRAG, a novel framework designed for open-question answering over spoken data. Our proposed approach fine-tunes a pre-trained speech encoder into a speech adapter fed into a frozen large language model (LLM)--based retrieval model. By aligning the embedding spaces of text and speech, our speech retriever directly retrieves audio passages from text-based queries, leveraging the retrieval capacity of the frozen text retriever. Our retrieval experiments on spoken question answering datasets show that direct speech retrieval does not degrade over the text-based baseline, and outperforms the cascaded systems using ASR. For generation, we use a speech language model (SLM) as a generator, conditioned on audio passages rather than transcripts. Without fine-tuning of the SLM, this approach outperforms cascaded text-based models when there is high WER in the transcripts.

Karel Mundnich, Xing Niu, Prashant Mathur et al. · amazon-science

Despite recent advancements in speech processing, zero-resource speech translation (ST) and automatic speech recognition (ASR) remain challenging problems. In this work, we propose to leverage a multilingual Large Language Model (LLM) to perform ST and ASR in languages for which the model has never seen paired audio-text data. We achieve this by using a pre-trained multilingual speech encoder, a multilingual LLM, and a lightweight adaptation module that maps the audio representations to the token embedding space of the LLM. We perform several experiments both in ST and ASR to understand how to best train the model and what data has the most impact on performance in previously unseen languages. In ST, our best model is capable to achieve BLEU scores over 23 in CoVoST2 for two previously unseen languages, while in ASR, we achieve WERs of up to 28.2\%. We finally show that the performance of our system is bounded by the ability of the LLM to output text in the desired language.

Raghuveer Peri, Sai Muralidhar Jayanthi, Srikanth Ronanki et al.

Integrated Speech and Large Language Models (SLMs) that can follow speech instructions and generate relevant text responses have gained popularity lately. However, the safety and robustness of these models remains largely unclear. In this work, we investigate the potential vulnerabilities of such instruction-following speech-language models to adversarial attacks and jailbreaking. Specifically, we design algorithms that can generate adversarial examples to jailbreak SLMs in both white-box and black-box attack settings without human involvement. Additionally, we propose countermeasures to thwart such jailbreaking attacks. Our models, trained on dialog data with speech instructions, achieve state-of-the-art performance on spoken question-answering task, scoring over 80% on both safety and helpfulness metrics. Despite safety guardrails, experiments on jailbreaking demonstrate the vulnerability of SLMs to adversarial perturbations and transfer attacks, with average attack success rates of 90% and 10% respectively when evaluated on a dataset of carefully designed harmful questions spanning 12 different toxic categories. However, we demonstrate that our proposed countermeasures reduce the attack success significantly.

Yufeng Du, Minyang Tian, Srikanth Ronanki et al.

Large language models (LLMs) often fail to scale their performance on long-context tasks performance in line with the context lengths they support. This gap is commonly attributed to retrieval failures -- the models' inability to identify relevant information in the long inputs. Accordingly, recent efforts often focus on evaluating and improving LLMs' retrieval performance: if retrieval is perfect, a model should, in principle, perform just as well on a long input as it does on a short one -- or should it? This paper presents findings that the answer to this question may be negative. Our systematic experiments across 5 open- and closed-source LLMs on math, question answering, and coding tasks reveal that, even when models can perfectly retrieve all relevant information, their performance still degrades substantially (13.9%--85%) as input length increases but remains well within the models' claimed lengths. This failure occurs even when the irrelevant tokens are replaced with minimally distracting whitespace, and, more surprisingly, when they are all masked and the models are forced to attend only to the relevant tokens. A similar performance drop is observed when all relevant evidence is placed immediately before the question. Our findings reveal a previously-unrealized limitation: the sheer length of the input alone can hurt LLM performance, independent of retrieval quality and without any distraction. They motivate our simple, model-agnostic mitigation strategy that transforms a long-context task into a short-context one by prompting the model to recite the retrieved evidence before attempting to solve the problem. On RULER, we observe a consistent improvement of GPT-4o up to 4% on an already strong baseline.

Goeric Huybrechts, Srikanth Ronanki, Sai Muralidhar Jayanthi et al.

The proliferation of multimodal Large Language Models has significantly advanced the ability to analyze and understand complex data inputs from different modalities. However, the processing of long documents remains under-explored, largely due to a lack of suitable benchmarks. To address this, we introduce Document Haystack, a comprehensive benchmark designed to evaluate the performance of Vision Language Models (VLMs) on long, visually complex documents. Document Haystack features documents ranging from 5 to 200 pages and strategically inserts pure text or multimodal text+image "needles" at various depths within the documents to challenge VLMs' retrieval capabilities. Comprising 400 document variants and a total of 8,250 questions, it is supported by an objective, automated evaluation framework. We detail the construction and characteristics of the Document Haystack dataset, present results from prominent VLMs and discuss potential research avenues in this area.

Woomin Song, Sai Muralidhar Jayanthi, Srikanth Ronanki et al.

As large language models increasingly gain popularity in real-world applications, processing extremely long contexts, often exceeding the model's pre-trained context limits, has emerged as a critical challenge. While existing approaches to efficient long-context processing show promise, recurrent compression-based methods struggle with information preservation, whereas random access approaches require substantial memory resources. We introduce REFORM, a novel inference framework that efficiently handles long contexts through a two-phase approach. First, it incrementally processes input chunks while maintaining a compressed KV cache, constructs cross-layer context embeddings, and utilizes early exit strategy for improved efficiency. Second, it identifies and gathers essential tokens via similarity matching and selectively recomputes the KV cache. Compared to baselines, REFORM achieves over 52% and 34% performance gains on RULER and BABILong respectively at 1M context length. It also outperforms baselines on Infinite-Bench, RepoEval, and MM-NIAH, demonstrating flexibility across diverse tasks and domains. Additionally, REFORM reduces inference time by 30% and peak memory usage by 5%, achieving both efficiency and superior performance.

Devang Kulshreshtha, Saket Dingliwal, Brady Houston et al.

Automatic Speech Recognition (ASR) traditionally assumes known domains, but adding data from a new domain raises concerns about computational inefficiencies linked to retraining models on both existing and new domains. Fine-tuning solely on new domain risks Catastrophic Forgetting (CF). To address this, Lifelong Learning (LLL) algorithms have been proposed for ASR. Prior research has explored techniques such as Elastic Weight Consolidation, Knowledge Distillation, and Replay, all of which necessitate either additional parameters or access to prior domain data. We propose Sequential Model Editing as a novel method to continually learn new domains in ASR systems. Different than previous methods, our approach does not necessitate access to prior datasets or the introduction of extra parameters. Our study demonstrates up to 15% Word Error Rate Reduction (WERR) over fine-tuning baseline, and superior efficiency over other LLL techniques on CommonVoice English multi-accent dataset.

Ashish Shenoy, Sravan Bodapati, Monica Sunkara et al.

Neural Language Models (NLM), when trained and evaluated with context spanning multiple utterances, have been shown to consistently outperform both conventional n-gram language models and NLMs that use limited context. In this paper, we investigate various techniques to incorporate turn based context history into both recurrent (LSTM) and Transformer-XL based NLMs. For recurrent based NLMs, we explore context carry over mechanism and feature based augmentation, where we incorporate other forms of contextual information such as bot response and system dialogue acts as classified by a Natural Language Understanding (NLU) model. To mitigate the sharp nearby, fuzzy far away problem with contextual NLM, we propose the use of attention layer over lexical metadata to improve feature based augmentation. Additionally, we adapt our contextual NLM towards user provided on-the-fly speech patterns by leveraging encodings from a large pre-trained masked language model and performing fusion with a Transformer-XL based NLM. We test our proposed models using N-best rescoring of ASR hypotheses of task-oriented dialogues and also evaluate on downstream NLU tasks such as intent classification and slot labeling. The best performing model shows a relative WER between 1.6% and 9.1% and a slot labeling F1 score improvement of 4% over non-contextual baselines.

Siddharth Dalmia, Yuzong Liu, Srikanth Ronanki et al.

We live in a world where 60% of the population can speak two or more languages fluently. Members of these communities constantly switch between languages when having a conversation. As automatic speech recognition (ASR) systems are being deployed to the real-world, there is a need for practical systems that can handle multiple languages both within an utterance or across utterances. In this paper, we present an end-to-end ASR system using a transformer-transducer model architecture for code-switched speech recognition. We propose three modifications over the vanilla model in order to handle various aspects of code-switching. First, we introduce two auxiliary loss functions to handle the low-resource scenario of code-switching. Second, we propose a novel mask-based training strategy with language ID information to improve the label encoder training towards intra-sentential code-switching. Finally, we propose a multi-label/multi-audio encoder structure to leverage the vast monolingual speech corpora towards code-switching. We demonstrate the efficacy of our proposed approaches on the SEAME dataset, a public Mandarin-English code-switching corpus, achieving a mixed error rate of 18.5% and 26.3% on test_man and test_sge sets respectively.

Monica Sunkara, Srikanth Ronanki, Dhanush Bekal et al.

In this work, we explore a multimodal semi-supervised learning approach for punctuation prediction by learning representations from large amounts of unlabelled audio and text data. Conventional approaches in speech processing typically use forced alignment to encoder per frame acoustic features to word level features and perform multimodal fusion of the resulting acoustic and lexical representations. As an alternative, we explore attention based multimodal fusion and compare its performance with forced alignment based fusion. Experiments conducted on the Fisher corpus show that our proposed approach achieves ~6-9% and ~3-4% absolute improvement (F1 score) over the baseline BLSTM model on reference transcripts and ASR outputs respectively. We further improve the model robustness to ASR errors by performing data augmentation with N-best lists which achieves up to an additional ~2-6% improvement on ASR outputs. We also demonstrate the effectiveness of semi-supervised learning approach by performing ablation study on various sizes of the corpus. When trained on 1 hour of speech and text data, the proposed model achieved ~9-18% absolute improvement over baseline model.

Monica Sunkara, Srikanth Ronanki, Kalpit Dixit et al.

Automatic speech recognition (ASR) systems in the medical domain that focus on transcribing clinical dictations and doctor-patient conversations often pose many challenges due to the complexity of the domain. ASR output typically undergoes automatic punctuation to enable users to speak naturally, without having to vocalise awkward and explicit punctuation commands, such as "period", "add comma" or "exclamation point", while truecasing enhances user readability and improves the performance of downstream NLP tasks. This paper proposes a conditional joint modeling framework for prediction of punctuation and truecasing using pretrained masked language models such as BERT, BioBERT and RoBERTa. We also present techniques for domain and task specific adaptation by fine-tuning masked language models with medical domain data. Finally, we improve the robustness of the model against common errors made in ASR by performing data augmentation. Experiments performed on dictation and conversational style corpora show that our proposed model achieves ~5% absolute improvement on ground truth text and ~10% improvement on ASR outputs over baseline models under F1 metric.

Xin Wang, Junichi Yamagishi, Massimiliano Todisco et al.

Automatic speaker verification (ASV) is one of the most natural and convenient means of biometric person recognition. Unfortunately, just like all other biometric systems, ASV is vulnerable to spoofing, also referred to as "presentation attacks." These vulnerabilities are generally unacceptable and call for spoofing countermeasures or "presentation attack detection" systems. In addition to impersonation, ASV systems are vulnerable to replay, speech synthesis, and voice conversion attacks. The ASVspoof 2019 edition is the first to consider all three spoofing attack types within a single challenge. While they originate from the same source database and same underlying protocol, they are explored in two specific use case scenarios. Spoofing attacks within a logical access (LA) scenario are generated with the latest speech synthesis and voice conversion technologies, including state-of-the-art neural acoustic and waveform model techniques. Replay spoofing attacks within a physical access (PA) scenario are generated through carefully controlled simulations that support much more revealing analysis than possible previously. Also new to the 2019 edition is the use of the tandem detection cost function metric, which reflects the impact of spoofing and countermeasures on the reliability of a fixed ASV system. This paper describes the database design, protocol, spoofing attack implementations, and baseline ASV and countermeasure results. It also describes a human assessment on spoofed data in logical access. It was demonstrated that the spoofing data in the ASVspoof 2019 database have varied degrees of perceived quality and similarity to the target speakers, including spoofed data that cannot be differentiated from bona-fide utterances even by human subjects.

Viacheslav Klimkov, Srikanth Ronanki, Jonas Rohnke et al.

We present a neural text-to-speech system for fine-grained prosody transfer from one speaker to another. Conventional approaches for end-to-end prosody transfer typically use either fixed-dimensional or variable-length prosody embedding via a secondary attention to encode the reference signal. However, when trained on a single-speaker dataset, the conventional prosody transfer systems are not robust enough to speaker variability, especially in the case of a reference signal coming from an unseen speaker. Therefore, we propose decoupling of the reference signal alignment from the overall system. For this purpose, we pre-compute phoneme-level time stamps and use them to aggregate prosodic features per phoneme, injecting them into a sequence-to-sequence text-to-speech system. We incorporate a variational auto-encoder to further enhance the latent representation of prosody embeddings. We show that our proposed approach is significantly more stable and achieves reliable prosody transplantation from an unseen speaker. We also propose a solution to the use case in which the transcription of the reference signal is absent. We evaluate all our proposed methods using both objective and subjective listening tests.

Nishant Prateek, Mateusz Łajszczak, Roberto Barra-Chicote et al.

Neural text-to-speech synthesis (NTTS) models have shown significant progress in generating high-quality speech, however they require a large quantity of training data. This makes creating models for multiple styles expensive and time-consuming. In this paper different styles of speech are analysed based on prosodic variations, from this a model is proposed to synthesise speech in the style of a newscaster, with just a few hours of supplementary data. We pose the problem of synthesising in a target style using limited data as that of creating a bi-style model that can synthesise both neutral-style and newscaster-style speech via a one-hot vector which factorises the two styles. We also propose conditioning the model on contextual word embeddings, and extensively evaluate it against neutral NTTS, and neutral concatenative-based synthesis. This model closes the gap in perceived style-appropriateness between natural recordings for newscaster-style of speech, and neutral speech synthesis by approximately two-thirds.

Javier Latorre, Jakub Lachowicz, Jaime Lorenzo-Trueba et al.

Recent speech synthesis systems based on sampling from autoregressive neural networks models can generate speech almost undistinguishable from human recordings. However, these models require large amounts of data. This paper shows that the lack of data from one speaker can be compensated with data from other speakers. The naturalness of Tacotron2-like models trained on a blend of 5k utterances from 7 speakers is better than that of speaker dependent models trained on 15k utterances, but in terms of stability multi-speaker models are always more stable. We also demonstrate that models mixing only 1250 utterances from a target speaker with 5k utterances from another 6 speakers can produce significantly better quality than state-of-the-art DNN-guided unit selection systems trained on more than 10 times the data from the target speaker.

Srikanth Ronanki, Oliver Watts, Simon King et al.

This paper proposes a new approach to duration modelling for statistical parametric speech synthesis in which a recurrent statistical model is trained to output a phone transition probability at each timestep (acoustic frame). Unlike conventional approaches to duration modelling -- which assume that duration distributions have a particular form (e.g., a Gaussian) and use the mean of that distribution for synthesis -- our approach can in principle model any distribution supported on the non-negative integers. Generation from this model can be performed in many ways; here we consider output generation based on the median predicted duration. The median is more typical (more probable) than the conventional mean duration, is robust to training-data irregularities, and enables incremental generation. Furthermore, a frame-level approach to duration prediction is consistent with a longer-term goal of modelling durations and acoustic features together. Results indicate that the proposed method is competitive with baseline approaches in approximating the median duration of held-out natural speech.

Srikanth Ronanki, Siva Reddy, Bajibabu Bollepalli et al.

Text-to-Speech synthesis in Indian languages has a seen lot of progress over the decade partly due to the annual Blizzard challenges. These systems assume the text to be written in Devanagari or Dravidian scripts which are nearly phonemic orthography scripts. However, the most common form of computer interaction among Indians is ASCII written transliterated text. Such text is generally noisy with many variations in spelling for the same word. In this paper we evaluate three approaches to synthesize speech from such noisy ASCII text: a naive Uni-Grapheme approach, a Multi-Grapheme approach, and a supervised Grapheme-to-Phoneme (G2P) approach. These methods first convert the ASCII text to a phonetic script, and then learn a Deep Neural Network to synthesize speech from that. We train and test our models on Blizzard Challenge datasets that were transliterated to ASCII using crowdsourcing. Our experiments on Hindi, Tamil and Telugu demonstrate that our models generate speech of competetive quality from ASCII text compared to the speech synthesized from the native scripts. All the accompanying transliterated datasets are released for public access.